1. Introduction
The rapid expansion of the low-altitude economy and the widespread adoption of unmanned aerial vehicles (UAVs) across military, public safety, logistics, and agricultural sectors have introduced unprecedented challenges to airspace security. The unauthorized and malicious flights of low-flying, slow, and small (LSS) unmanned aerial vehicles pose increasingly severe threats to national security and public order. Effective detection and continuous tracking of these LSS unmanned aerial vehicles targets constitute the foundational technical pillar of counter-UAV systems.
Various sensing modalities exist for UAV detection, including radar, radio frequency (RF) analysis, infrared thermography, acoustic sensing, and visible-light vision. Radar systems, while robust in all-weather conditions, suffer from limited accuracy for small UAV targets and high deployment costs. RF detection depends on active communication links from the drone, which fails against autonomous or communication-silent vehicles. Thermal and acoustic methods are constrained by environmental noise, range limitations, and resolution constraints. In contrast, vision-based detection using visible-light cameras offers distinct advantages: low cost, rich information content, easy deployment, and natural integration with existing surveillance infrastructure.
However, LSS unmanned aerial vehicles present formidable challenges to general vision algorithms. At typical operational distances, these targets occupy only a minuscule fraction of the image, often appearing as just 10 to 30 pixels in size. They exhibit low contrast against cluttered backgrounds, possess weak textural features, and undergo rapid, agile maneuvers. Occlusions, motion blur, and sudden illumination changes further exacerbate the difficulty. Traditional hand-crafted feature methods prove insufficient under such complex conditions, failing to meet the combined demands of accuracy, robustness, and real-time processing required by practical counter-UAV operations.
The advent of deep learning, particularly convolutional neural networks and Vision Transformers, has driven significant progress in object detection and tracking. The YOLO family of detectors has established itself as the dominant approach for real-time detection, while tracking algorithms such as SORT, DeepSORT, and ByteTrack have continuously refined the data association and trajectory management processes. Nevertheless, direct application of these generic algorithms to LSS unmanned aerial vehicles scenarios still suffers from critical shortcomings: high miss rates for small targets, insufficient tracking stability under maneuver, and a persistent tension between computational complexity and real-time performance.
This thesis aims to systematically address these challenges through a three-layer research framework. First, I develop an enhanced detection algorithm, YOLO11-DWL, which incorporates structural re-parameterization, wavelet-based pooling, and an advanced detection head to improve small-target feature extraction and context modeling. Second, I propose an improved tracking algorithm, S-Bot-SORT, which replaces the standard IoU metric with a comprehensive SIoU loss to enhance data association stability. Finally, I design and implement an embedded detection and tracking system, integrating these algorithms on an NVIDIA Jetson platform for real-world deployment. The remainder of this paper is organized as follows: Section 2 reviews the relevant theoretical foundations; Section 3 details the improved detection algorithm; Section 4 presents the tracking algorithm enhancement; Section 5 describes the embedded system design and evaluation; Section 6 concludes the paper.
2. Theoretical Foundations
2.1 Deep Learning for Object Detection
Object detection identifies and localizes all objects of specified classes within a given image, producing detection outputs as a set of tuples. Modern detection methods are broadly classified into two-stage detectors and one-stage detectors. Two-stage methods, exemplified by Faster R-CNN, first generate region proposals and then classify and refine them. One-stage methods such as the YOLO series directly regress object locations and class probabilities from feature maps, achieving an excellent balance between speed and accuracy, which makes them the preferred choice for real-time LSS unmanned aerial vehicles detection.
The core feature extractor in modern detectors is the convolutional neural network (CNN). Given an input feature map $X \in \mathbb{R}^{H \times W \times C_{in}}$ and a convolutional kernel $K \in \mathbb{R}^{k_h \times k_w \times C_{in} \times C_{out}}$, the convolution output $Y$ at position $(i, j)$ and channel $o$ is calculated as follows:
$$Y_{i,j,o} = \sum_{c=1}^{C_{in}} \sum_{v=1}^{k_w} \sum_{u=1}^{k_h} K_{u,v,c,o} \cdot X_{i+s u, j+s v, c} + b_o \quad (2-1)$$
where $s$ is the stride. The receptive field $r_l$ of a network with $L$ convolution layers follows the recursive relation:
$$r_l = r_{l-1} + (k_l – 1) \prod_{i=1}^{l-1} s_i \quad (2-2)$$
where $k_l$ is the kernel size of layer $l$ and $s_i$ is the stride of preceding layers. For small targets, the receptive field must match the target scale. If a target has pixel dimensions $h_t \times w_t$, the effective receptive field $r$ should satisfy $r \geq \alpha \cdot \max(h_t, w_t)$ with $\alpha \in [2,4]$ for sufficient context modeling. Continuous downsampling operations, while expanding the receptive field, reduce spatial resolution and can degrade small-target features into a single feature point in deep layers.
Structural re-parameterization offers a principled approach to enhance feature extraction without increasing inference cost. The mathematical essence is the equivalent superposition of multiple linear operators. If a module has $M$ parallel convolutional branches, the sum of their outputs can be expressed as a single merged convolution:
$$\mathcal{Y} = \sum_{m=1}^{M} (K^{(m)} * X + b^{(m)}) = \left(\sum_{m=1}^{M} K^{(m)}\right) * X + \sum_{m=1}^{M} b^{(m)} \quad (2-3)$$
When branches are followed by batch normalization, the equivalent kernel and bias for branch $m$ become:
$$\tilde{K}^{(m)} = \frac{\gamma^{(m)}}{\sigma^{(m)}} K^{(m)}, \quad \tilde{b}^{(m)} = \frac{\gamma^{(m)}}{\sigma^{(m)}} (b^{(m)} – \mu^{(m)}) + \beta^{(m)} \quad (2-4)$$
where $\mu^{(m)}$ and $\sigma^{(m)}$ are the batch statistics, and $\gamma^{(m)}$ and $\beta^{(m)}$ are the learnable affine parameters. This technique enables multi-branch training for better gradient flow and implicit regularization, while ensuring single-branch inference efficiency.
2.2 Deep Learning for Object Tracking
The dominant paradigm in multi-object tracking is the tracking-by-detection (TBD) approach, which decouples the problem into three independent sub-problems: detection, motion prediction, and data association. The core mathematical formulation of data association is a constrained linear assignment problem:
$$\min_{X} \sum_{i,j} C_{ij} X_{ij} \quad (2-5)$$
where $X_{ij} \in \{0,1\}$ indicates whether track $i$ is matched to detection $j$, and $C_{ij}$ is the cost matrix. The standard IoU metric, defined as:
$$IoU(\hat{b}, b) = \frac{|\hat{b} \cap b|}{|\hat{b} \cup b|} \quad (2-6)$$
serves as the core metric for motion similarity. However, for LSS unmanned aerial vehicles tracking, the standard IoU exhibits critical limitations. For a target of side length $s$ pixels with a center offset $\delta$, the IoU decays approximately quadratically:
$$IoU \approx 1 – \frac{2\delta}{s} + O\left(\frac{\delta^2}{s^2}\right) \quad (2-7)$$
When $\delta / s > 0.5$, the IoU approaches zero. For target sizes of only 10 to 30 pixels, Kalman filter prediction errors under agile maneuvers often exceed half the target size, causing legitimate track-detection pairs to be misjudged as non-matching, leading to track fragmentation and identity switches. Additionally, standard IoU lacks directional and geometric constraints, failing to distinguish between reasonable scale changes and erroneous lateral drifts.
2.3 Characteristics of UAV Targets in Detection and Tracking
LSS unmanned aerial vehicles targets exhibit several unique characteristics that challenge general vision algorithms:
Low resolution and small target scale: The pixel occupancy of LSS UAV targets is typically below 0.1% of the image area. According to the Nyquist-Shannon sampling theorem, when the target spatial frequency exceeds the image sampling frequency, high-frequency details are lost or aliased. Continuous downsampling by a factor of 2 further degrades the effective information of small targets layer by layer.
Complex background interference: UAV targets often appear against backgrounds with sky, clouds, buildings, and terrain. High background similarity makes it difficult for classifiers to distinguish the target from distractors, increasing false positive rates.
Rapid and non-linear motion: UAVs exhibit agile maneuvers, rapid turns, and non-uniform motion patterns, which violate the constant-velocity assumption of standard Kalman filters, leading to significant prediction errors and association failures.
Occlusion and target loss recovery: When targets are occluded by clouds or structures, the detector output is missing, and trajectories enter an unmatched state. The prediction error accumulates over time, and upon target reappearance, the IoU between the extrapolated box and the new detection may be too low to recover the trajectory.
3. Detection Algorithm for LSS UAVs
3.1 Baseline Algorithm Analysis
YOLO11 serves as the baseline detection framework due to its excellent balance between inference speed and accuracy. The architecture follows the Backbone-Neck-Head paradigm. The backbone network employs the C3k2 module, which replaces the C2f module of YOLOv8, achieving more efficient feature extraction through a cross-stage partial connection mechanism. The spatial pyramid pooling fast (SPPF) module captures multi-scale contextual information, followed by a cross-stage partial spatial attention module that enhances key spatial features.
The Neck network adopts a bidirectional feature pyramid network structure, enabling efficient multi-scale feature fusion through top-down and bottom-up propagation paths. The detection head employs a decoupled design, separating classification and regression tasks into independent branches, and incorporates depthwise separable convolutions to reduce parameter count and computational complexity.
3.2 Proposed YOLO11-DWL Algorithm
Despite the strong performance of YOLO11 on general object detection benchmarks, I identify several critical defects when applying it to LSS unmanned aerial vehicles detection. First, the continuous downsampling in the backbone causes severe small-target feature degradation; after 32× downsampling, discriminative information for targets below 32 pixels nearly disappears in the final feature maps. Second, the fixed-weight fusion of multi-scale features cannot adapt to datasets heavily biased toward small objects. Third, the standard detection head lacks contextual sensitivity, leading to poor differentiation between low-contrast UAV targets and cluttered backgrounds.
To overcome these limitations, I propose the YOLO11-DWL algorithm, which integrates three novel components: a structural re-parameterization module, a wavelet-based pooling mechanism, and an advanced context-aware detection head.
3.2.1 Structural Re-parameterization Module
The original C3k2 module is replaced with an enhanced version, C3k2-DRB, which incorporates a dilated-residual dual-branch architecture. The forward propagation process is described by:
$$\mathcal{Y} = \mathcal{C}_{1\times1}^{Concat}[\mathcal{P}_1, \mathcal{C}_{DB}(\mathcal{P}_2)] \quad (3-1)$$
where $\mathcal{P}_1 = \mathcal{S}_c(x)$ is the channel-sliced half-dimensional residual feature, and the DRB transformation $\mathcal{C}_{DB}(\cdot)$ is defined as:
$$\mathcal{C}_{DB}(x) = \sigma(\mathcal{W}_{1\times1}[x + \mathcal{W}_{3\times3}^{res}(x) \| \mathcal{W}_{3\times3}^{dil}(x; r=2)]) \quad (3-2)$$
Here, $\|$ denotes channel concatenation, and $\mathcal{W}_{3\times3}^{dil}(\cdot; r=2)$ is a dilated convolution with dilation rate $r=2$, which expands the effective receptive field to 5×5 while maintaining $9C_{in}C_{out}$ parameters. The receptive field enhancement can be quantified as:
$$RF_{eff}^{(l)} = RF_{eff}^{(l-1)} + (k-1) \times r \times \prod_{i=1}^{l-1} s_i \quad (3-3)$$
At the P3 level, the DRB increases the receptive field from 15×15 to 27×27 pixels, an improvement of 44%, effectively capturing the long-range motion trajectory and background contrast cues of low-altitude UAVs. The gradient flow through the residual path is:
$$\frac{\partial \mathcal{L}}{\partial x} = \frac{\partial \mathcal{L}}{\partial \mathcal{Y}} \left(1 + \frac{\partial \mathcal{W}_{3\times3}^{res}(x)}{\partial x}\right) \quad (3-4)$$
This additional term effectively alleviates the vanishing gradient problem in deep networks. The DRB strictly follows a parameter-neutral design principle:
$$\Delta_{Params} = Params_{C3K2-DRB} – Params_{C3k2} \equiv 0 \quad (3-5)$$
3.2.2 Wavelet-based Pooling Convolution
Standard convolution with stride 2 acts as a low-pass filter that smooths the high-frequency details of small targets. To address this, I propose the WaveletPool parameter-free downsampling mechanism based on discrete wavelet transform (DWT). The wavelet basis functions construct a multi-resolution analysis framework:
$$\varphi_{j,k}(x,y) = 2^j \varphi(2^j x – k_x, 2^j y – k_y) \quad (3-6)$$
The WaveletPool employs a parameter-free Haar wavelet kernel, and the decomposition of each 2×2 block is matrix-formulated as:
$$\begin{bmatrix} Y_{i,j}^{LL} \\ Y_{i,j}^{LH} \\ Y_{i,j}^{HL} \\ Y_{i,j}^{HH} \end{bmatrix} = \frac{1}{4} \begin{bmatrix} +1 & +1 & +1 & +1 \\ +1 & -1 & +1 & -1 \\ +1 & +1 & -1 & -1 \\ +1 & -1 & -1 & +1 \end{bmatrix} \begin{bmatrix} X_{2i,2j} \\ X_{2i+1,2j} \\ X_{2i,2j+1} \\ X_{2i+1,2j+1} \end{bmatrix} \quad (3-7)$$
where $Y^{LL}$ is the approximation coefficient (low-frequency component) that preserves global structure, and $Y^{LH}, Y^{HL}, Y^{HH}$ are the horizontal, vertical, and diagonal detail coefficients (high-frequency components). The WaveletPool outputs $Y^{LL}$ as the downsampled result while injecting the high-frequency components into subsequent C3k2 modules via cross-layer connections, achieving lossless information preservation.
3.2.3 Large-Small Context Sampling and Boundary-Discriminative Detection Head
The improved detection head, termed LSCSBD, introduces three key mechanisms. First, a scale-adaptive weighted fusion mechanism employs learnable weights for multi-scale feature integration:
$$\hat{F} = \frac{\sum_{i=3}^{5} \text{ReLU}(w_i) \cdot \text{Resample}(\tilde{F}_i)}{\sum_{i=3}^{5} \text{ReLU}(w_i) + \epsilon} \quad (3-8)$$
where $w_i \geq 0$ are learnable scalar weights and Resample denotes resizing to the target resolution. Second, a context-sensitive attention module concatenates channel and spatial attention. The channel attention is computed as:
$$z_c = \text{GAP}(F) \in \mathbb{R}^C, \quad s_c = \sigma(W_2 \delta(W_1 z_c)), \quad F’ = s_c \odot F \quad (3-9)$$
where GAP is the global average pooling operation. The spatial attention then operates on the channel-calibrated features:
$$s_s = \sigma(f^{7\times7}([\text{AvgPool}(F’); \text{MaxPool}(F’)])), \quad F” = s_s \odot F’ \quad (3-10)$$
Third, the total loss function is composed of three parts, where the regression loss uses CIoU, and the classification and confidence losses use Focal Loss:
$$\mathcal{L} = \lambda_{box} \mathcal{L}_{CIoU} + \lambda_{obj} \mathcal{L}_{Focal} + \lambda_{cls} \mathcal{L}_{Focal} \quad (3-11)$$
The CIoU loss is defined as:
$$\mathcal{L}_{CIoU} = 1 – IoU + \frac{d^2}{c^2} + \alpha v \quad (3-12)$$
where $d$ is the Euclidean distance between the centers of the predicted and ground-truth boxes, $c$ is the diagonal length of the smallest enclosing box, $v$ is the aspect ratio consistency term, and $\alpha$ is the adaptive scaling factor. The Focal Loss is given by:
$$FL(p_t) = -\alpha (1-p_t)^\gamma \log(p_t) \quad (3-13)$$
with typical hyperparameters $\gamma=2$ and $\alpha=0.25$, which effectively mitigates the extreme foreground-background class imbalance in small target detection.
3.3 Experimental Setup and Results
3.3.1 Experimental Configuration
All experiments were conducted on an NVIDIA GeForce RTX 5060Ti GPU. The models were trained for 300 epochs with an initial learning rate of 0.01, batch size of 32, momentum of 0.937, and weight decay of 0.0005. A 3-epoch warm-up training was employed to stabilize early training. The evaluation metrics include detection precision, recall, mAP@0.5, and model parameter count.
| Parameter | Value |
|---|---|
| Input image size | 640×640 |
| Batch size | 32 |
| Epochs | 300 |
| Initial learning rate | 0.01 |
| Momentum | 0.937 |
| Weight decay | 0.0005 |
| Warm-up epochs | 3 |
3.3.2 Ablation Experiments
To verify the effectiveness of each improvement module, I conducted ablation experiments on a self-constructed dataset containing 6,469 images collected in real outdoor scenes, covering beach, urban, suburban, and forest backgrounds. The results are presented in the following table.
| Model | P (%) | R (%) | mAP@0.5 (%) | FPS | Params (M) |
|---|---|---|---|---|---|
| YOLO11 | 92.6 | 75.2 | 83.7 | 269.5 | 2.58 |
| YOLO11-DRB | 92.7 | 76.1 | 84.0 | 213.4 | 2.44 |
| YOLO11-WaveletPool | 92.9 | 76.0 | 83.8 | 250.0 | 2.17 |
| YOLO11-LSCSBD | 92.9 | 77.2 | 84.9 | 238.3 | 2.46 |
| YOLO11-DWL | 92.9 | 78.4 | 85.8 | 196.9 | 1.90 |
From the ablation results, the structural re-parameterization module improves recall by 0.9% and mAP by 0.3%, demonstrating that the dilated convolution effectively expands the receptive field and captures more contextual information. The WaveletPool reduces parameters by 15.8% while achieving comparable precision, confirming the effectiveness of wavelet pooling in preserving high-frequency information. The LSCSBD detection head brings the most significant improvement, increasing recall by 2.0% and mAP by 1.2%, validating the multi-scale weighted fusion and attention mechanisms. The complete YOLO11-DWL model achieves the best overall performance with the lowest parameter count (1.90M), an improvement of 26.4% parameter reduction, while improving mAP@0.5 by 2.1 percentage points compared to the baseline.
3.3.3 Generalization Experiments on Public Datasets
To evaluate the generalization capability of the proposed method, I further validated the model on two public benchmark datasets: the DUT Anti-UAV detection dataset and the TIB-NET dataset.
| Dataset | Model | P (%) | R (%) | mAP@0.5 (%) | Params (M) |
|---|---|---|---|---|---|
| DUT Anti-UAV | YOLO11 | 92.6 | 83.1 | 88.6 | 2.58 |
| YOLO11-DWL | 92.7 | 83.1 | 89.3 | 1.90 | |
| TIB-NET | YOLO11 | 87.3 | 78.3 | 83.8 | 2.58 |
| YOLO11-DWL | 86.0 | 79.8 | 86.6 | 1.90 |
On the DUT Anti-UAV dataset, the proposed YOLO11-DWL achieves a mAP@0.5 of 89.3%, outperforming the baseline by 0.7 percentage points. On the TIB-NET dataset, the improvement is even more substantial, with mAP@0.5 increasing from 83.8% to 86.6%, a significant gain of 2.8 percentage points. The recall rate also improved from 78.3% to 79.8%, indicating that the WaveletPool preserves high-frequency information for effective separation of targets from strong background interference. These results demonstrate the strong generalization capability of the “structural re-parameterization + wavelet pooling + context-enhanced detection head” improvement combination across diverse scenarios.
4. Tracking Algorithm for LSS UAVs
4.1 Baseline Algorithm
I select Bot-SORT as the baseline tracking framework due to its excellent performance in multi-object tracking. Bot-SORT builds upon the tracking-by-detection paradigm with three key innovations. First, it employs an enhanced Kalman filter with an 8-dimensional state vector that explicitly models width and height variations independently:
$$X_t = [x_c, y_c, w, h, v_x, v_y, v_w, v_h]^T \in \mathbb{R}^8 \quad (4-1)$$
where $(x_c, y_c)$ is the center coordinate, $(w, h)$ represents the bounding box width and height, and $(\dot{x}_c, \dot{y}_c, \dot{w}, \dot{h})$ are the corresponding first-order velocity terms. This design decouples target motion into translational motion $(x_c, y_c, v_x, v_y)$ and deformation motion $(w, h, v_w, v_h)$, which aligns with the projection laws of UAV targets in the image plane.
Second, Bot-SORT introduces a camera motion compensation (GMC) module that estimates the global motion between consecutive frames using sparse optical flow or ORB feature points. The camera motion is approximated by a 2D affine transformation:
$$\begin{bmatrix} x’ \\ y’ \end{bmatrix} = M \begin{bmatrix} x \\ y \end{bmatrix} + T \quad (4-2)$$
where $M \in \mathbb{R}^{2\times2}$ contains rotation and scale, and $T \in \mathbb{R}^{2\times1}$ is the translation vector. The RANSAC algorithm robustly estimates the affine matrix $A_{k-1}^{k}$ by minimizing:
$$A_{k-1}^{k} = \arg\min_A \sum_{i=1}^{N} |p_i^k – A p_i^{k-1}|^2 \quad (4-3)$$
Third, Bot-SORT integrates IoU and Re-ID features in a hierarchical association strategy. The appearance features are extracted using a FastReID backbone network with the generalized mean (GeM) pooling:
$$f = \left(\frac{1}{|X|} \sum_{x \in X} x^p\right)^{1/p} \quad (4-4)$$
where $p$ is a learnable parameter that adaptively balances between average pooling ($p=1$) and max pooling ($p \to \infty$). The appearance feature is updated by exponential moving average for matched tracks:
$$e_i^k = \alpha e_i^{k-1} + (1-\alpha) f_j^k \quad (4-5)$$
The final association cost integrates both IoU and cosine distance:
$$\hat{d}_{i,j}^{cos} = \begin{cases} 0.5 \cdot d_{i,j}^{cos} & \text{if } (d_{i,j}^{cos} < \theta_{emb}) \wedge (d_{i,j}^{IoU} < \theta_{IoU}) \\ 1 & \text{otherwise} \end{cases} \quad (4-6)$$
The conservative strategy ensures that the matching cost is reduced only when both spatial and appearance similarities are sufficiently high, avoiding mis-associations caused by relying on a single metric.
4.2 Limitations of Standard IoU for LSS UAV Tracking
In the data association module of Bot-SORT, IoU is used as the spatial similarity metric, forming the basis of the cost matrix. However, for LSS unmanned aerial vehicles tracking, standard IoU exhibits three critical deficiencies. First, the target pixel area is extremely small, and slight displacements cause IoU to drop sharply. When a target is occluded, the detection output is missing, and the IoU between the predicted trajectory box and the true detection remains at zero, leading to trajectory loss. Second, during rapid scale changes as the UAV approaches the camera, the bounding box area can change by over 100 times. For a 20×20 pixel target, a 1-pixel center offset causes IoU to drop by 0.12, while for a 100×100 pixel target, the same offset causes only a 0.03 drop. This nonlinear sensitivity difference makes it difficult to set a unified IoU threshold. Third, standard IoU lacks directional and geometric prior information, failing to distinguish between center-aligned scale changes and offset center positions.
4.3 Proposed S-Bot-SORT Algorithm
To overcome these limitations, I propose the S-Bot-SORT algorithm, which replaces standard IoU with the SIoU metric in the data association stage. SIoU is a comprehensive regression loss composed of four parts: angle loss, distance loss, shape loss, and IoU loss.
Angle loss: This component penalizes the deviation of the center point connection direction from the principal axis of the prediction box:
$$\Lambda(\alpha, \beta) = \begin{cases} 1 – 2\sin^2\left(\arcsin(x) – \frac{\pi}{4}\right) & \text{if } |\alpha| \leq \frac{\pi}{4} \\ 1 – 2\sin^2\left(\arcsin(x) – \frac{3\pi}{4}\right) & \text{if } |\alpha| > \frac{\pi}{4} \end{cases} \quad (4-7)$$
where the parameter $x$ is calculated as:
$$x = \frac{\sqrt{(b_{x_i} – \hat{b}_{x_j})^2 + (b_{y_i} – \hat{b}_{y_j})^2}}{C_v} \quad (4-8)$$
Distance loss: Building upon the angle loss, the distance loss further penalizes the Euclidean distance between center points:
$$\Delta(\rho, \gamma) = 1 – \exp(-\gamma \cdot \rho^2) \quad (4-9)$$
where $\rho$ is the normalized center distance and $\gamma$ is the sensitivity coefficient that couples with the angle loss: $\gamma = 2 – \Lambda(\alpha, \beta)$.
Shape loss: The shape loss penalizes aspect ratio differences to address bounding box deformation caused by UAV pitch angle changes:
$$\Omega = \left(1 – e^{-\left(\frac{|w_i – w_j|}{\max(w_i, w_j)}\right)^\theta}\right) + \left(1 – e^{-\left(\frac{|h_i – h_j|}{\max(h_i, h_j)}\right)^\theta}\right) \quad (4-10)$$
where $\theta = 0.5$ is the shape sensitivity index, which weakens the shape constraint to avoid excessive punishment of reasonable scaling.
Comprehensive SIoU cost function: The final SIoU similarity is obtained by converting from the negative loss:
$$SIoU(b_i, \hat{b}_j) = 1 – \left[\underbrace{(1-IoU)}_{\text{IoU loss}} + \underbrace{\frac{\Lambda}{2}}_{\text{angle loss}} + \underbrace{\frac{\Delta}{4}}_{\text{distance loss}} + \underbrace{\frac{\Omega}{8}}_{\text{shape loss}}\right] \quad (4-11)$$
The denominator design reflects the priority: IoU loss weight of 1, angle loss of 0.5, distance loss of 0.25, and shape loss of 0.125, consistent with the principle of decreasing geometric constraint importance. The SIoU range is $[-0.125, 1]$, with negative values appearing when IoU is extremely low and geometric deviation is large, serving as a hard rejection signal.
4.4 Experimental Results and Analysis
4.4.1 Evaluation Metrics
I employ four complementary tracking evaluation metrics. HOTA comprehensively measures detection, association, and localization accuracy. MOTA focuses on the overall detection and association performance. MOTP specifically evaluates bounding box localization accuracy. IDF1 reflects identity preservation capability and long-term tracking continuity.
The MOTA metric is defined as:
$$MOTA = 1 – \frac{\sum_{t=1}^{T} (FN_t + FP_t + IDSW_t)}{\sum_{t=1}^{T} GT_t} \quad (4-12)$$
where $FN_t$ is the number of false negatives, $FP_t$ is the number of false positives, and $IDSW_t$ is the number of ID switches in frame $t$. The IDF1 metric is calculated as:
$$IDF1 = \frac{2 \cdot IDTP}{2 \cdot IDTP + IDFP + IDFN} \quad (4-13)$$
4.4.2 Joint Detection-Tracking Experiments
First, I compared different YOLO detector versions combined with Bot-SORT on the DUT Anti-UAV tracking subset. The results are presented in the following table.
| Detector | Tracker | HOTA | MOTA | MOTP | IDF1 |
|---|---|---|---|---|---|
| YOLOv5 | Bot-SORT | 52.676 | 51.039 | 76.820 | 74.816 |
| YOLOv8 | Bot-SORT | 52.945 | 52.154 | 76.966 | 75.112 |
| YOLO11 | Bot-SORT | 53.788 | 51.919 | 77.403 | 75.960 |
YOLO11 combined with Bot-SORT achieves the best HOTA of 53.788%, improving by 1.112% and 0.843% over YOLOv5 and YOLOv8, respectively. The MOTP and IDF1 values of 77.403% and 75.960% are also superior to the other configurations, indicating that YOLO11’s enhanced feature extraction and decoupled detection head produce more accurate and reliable detection results for the tracking system.
To systematically evaluate the independent contributions and synergistic effects of the detection improvements and tracking improvements, I designed four progressive comparison experiments as follows.
| Group | Detector | Tracker | HOTA | MOTA | MOTP | IDF1 |
|---|---|---|---|---|---|---|
| A | YOLO11 | Bot-SORT | 53.788 | 51.919 | 77.403 | 75.960 |
| B | YOLO11-DWL | Bot-SORT | 54.190 | 53.032 | 77.485 | 76.516 |
| C | YOLO11 | S-Bot-SORT | 54.244 | 52.330 | 77.678 | 76.165 |
| D | YOLO11-DWL | S-Bot-SORT | 54.257 | 53.129 | 77.783 | 76.564 |
The comparison between Group A and Group B reveals that replacing the detector with YOLO11-DWL while keeping the tracker unchanged results in HOTA improvement of 0.402%, MOTA improvement of 1.113%, MOTP improvement of 0.082%, and IDF1 improvement of 0.556%. This result clearly validates the effectiveness of the YOLO11-DWL detection improvements. The enhanced detection accuracy directly reduces miss rates and false positives, manifesting as a significant MOTA gain in the tracking stage. The detection improvements serve as the dominant factor because, within the tracking-by-detection paradigm, the observation $z_k$ in the Kalman filter update equation $x_{k|k-1} = x_{k|k-1} + K_k(z_k – H x_{k|k-1})$ contains detection error $\epsilon_t$, which propagates through the Kalman gain into all subsequent state estimations. Therefore, improving detection accuracy directly reduces the error source.
The comparison between Group C and Group A shows that the SIoU-based tracking improvement provides limited gains. HOTA improves by 0.456%, MOTA by 0.411%, and IDF1 by 0.205%. This phenomenon indicates that detection quality is the primary bottleneck for tracking performance in the DUT Anti-UAV dataset. The SIoU improvement plays a role in compensating for shortcomings when the detector is weaker, but once the detector is optimized, the association module is no longer the main bottleneck.
The complete scheme (Group D) achieves the best overall performance: HOTA improvement of 0.469%, MOTA improvement of 1.210%, MOTP improvement of 0.380%, and IDF1 improvement of 0.604% over the baseline. This demonstrates that detection accuracy is the core bottleneck for LSS unmanned aerial vehicles tracking, and data association strategy optimization can only fully demonstrate its utility when detection accuracy reaches a certain level.
5. Embedded Detection and Tracking System
5.1 System Architecture
To validate the engineering practicality of the proposed algorithms, I designed and implemented an embedded detection and tracking system for LSS unmanned aerial vehicles. The system adopts a modular, layered architecture comprising three levels: hardware perception layer, algorithm processing layer, and application interaction layer.
The hardware perception layer uses the NN-Q40XT50 dual-spectrum anti-shake spherical photoelectric turntable as the core sensing device. This device provides a 1/1.8-inch CMOS sensor with 1920×1080 resolution, achieving high-quality imaging under various lighting conditions. The high-precision servo system provides horizontal positioning accuracy of 0.005° and vertical rotation range from +90° to -90°, satisfying the requirements for capturing LSS unmanned aerial vehicles targets at various flight altitudes.
The algorithm processing layer is deployed on the NVIDIA Jetson Xavier NX platform, which provides up to 21 TOPS of AI inference performance with a power consumption of only 15W. This platform was selected after a comprehensive comparison with alternatives, as detailed in the following table.
| Platform | TOPS | GPU Cores | CPU Cores | Memory (GB) | Power (W) |
|---|---|---|---|---|---|
| Jetson Nano | 0.5 | 128 | 4 | 4 | 5-10 |
| Jetson TX2 | 1.3 | 256 | 6 | 8 | 7.5-15 |
| Jetson Xavier NX | 21 | 384+48 | 6 | 8/16 | 10-15 |
| Jetson AGX Xavier | 32 | 512+64 | 8 | 16 | 10-30 |
The Jetson Xavier NX offers the best balance between AI inference capability, power consumption, and interface richness. Its native support for CUDA, TensorRT, and PyTorch enables efficient deployment of deep learning models through the “PyTorch → ONNX → TensorRT” optimization pipeline.
5.2 Software System Design
The software system is built with a modular design philosophy using Python, PyQt5, OpenCV, and TensorRT. The system receives the RTSP video stream from the photoelectric turntable through a gigabit Ethernet interface on the Jetson Xavier NX platform. The preprocessing pipeline includes resizing images to 640×640, color space conversion from BGR to RGB, and pixel value normalization. An adaptive histogram equalization step is incorporated as an optional pre-processing step to handle challenging outdoor illumination conditions.
For efficient inference on the embedded platform, I employed TensorRT to optimize the trained PyTorch model. The conversion process involves three stages: first, the best.pt model is converted to ONNX format; second, the ONNX model is optimized using the TensorRT conversion tool, which performs graph optimization, layer fusion, and precision calibration; finally, a highly optimized engine file is generated for inference.
The S-Bot-SORT tracker is integrated as a Python library, sharing preprocessed image data with the detector for seamless inter-operation. The PyQt5-based graphical user interface provides a visual display of the video stream with overlaid detection boxes, tracking trajectories, confidence scores, and track IDs. A control panel allows operators to configure model loading, detection/tracking mode switching, confidence thresholds, and data recording functions.
5.3 System Performance Testing
I conducted comprehensive system tests in real outdoor scenarios using a DJI Mavic AIR 2 as the target unmanned aerial vehicle. The flight speed was limited to 50 m/s and the range to 100m to replicate the characteristics of LSS unmanned aerial vehicles. Tests were conducted in three distinct environmental conditions: open coastline, urban suburb, and complex building environments.
| Scenario | Detection and Tracking Accuracy (%) | FPS (frames per second) |
|---|---|---|
| Open coast | 70 | 51 |
| Urban suburb | 59 | 49 |
| Complex building | 57 | 48 |
The system achieves stable processing frame rates above 30 FPS in all tested scenarios, satisfying the real-time detection and tracking requirements. The open coast scenario yielded the best performance with a detection and tracking accuracy of 70%, where the system successfully distinguished UAV targets from distant reef edges and ocean waves while maintaining stable target tracking IDs even during brief occlusions and scale changes. In the urban suburb scenario, the system correctly identified UAV targets against complex backgrounds, including distinguishing them from visually similar objects such as distant streetlights. The complex building scenario presented the greatest challenge due to the presence of dark LED screens and glass curtain walls with strong textural interference, yet the system still maintained acceptable performance, achieving 57% accuracy. The experimental results demonstrate that the system is capable of stably identifying small UAV targets within a 100-meter range under diverse real-world conditions, confirming the engineering practicality and environmental adaptability of the proposed solution.
6. Conclusion and Future Work
6.1 Conclusion
This thesis systematically addressed the key challenges in detection and tracking of low-flying, slow, and small unmanned aerial vehicles targets. Through the design of algorithm improvements, rigorous experimental validation, and engineering integration, the main contributions are summarized as follows.
First, I proposed the YOLO11-DWL detection algorithm that achieves significant precision and lightweight optimization through three key innovations. The structural re-parameterization module expands the effective receptive field to enhance contextual modeling of small targets. The WaveletPool parameter-free downsampling mechanism preserves high-frequency details and reduces model parameters by 26.4%. The LSCSBD detection head enhances multi-scale feature fusion with learnable weights and attention mechanisms. The complete model achieves an mAP@0.5 of 85.8% on the self-constructed dataset, improving by 2.1 percentage points over the baseline while reducing parameters to 1.90M. Validation on two public benchmark datasets confirmed the algorithm’s excellent generalization capability.
Second, I proposed the S-Bot-SORT tracking algorithm that replaces the standard IoU metric with a four-dimensional SIoU weighted fusion to enhance data association stability. Through four progressive comparison experiments, I comprehensively analyzed the independent contributions and synergistic effects of detection improvements and tracking improvements. The key finding is that detection accuracy is the primary bottleneck for LSS unmanned aerial vehicles tracking performance, with YOLO11-DWL detection improvements yielding a MOTA gain of 1.113% compared to the limited contribution of tracking improvements of 0.097% on top of the optimal detector.
Third, I designed and implemented a complete embedded detection and tracking system. The system integrates the YOLO11-DWL and S-Bot-SORT algorithms on an NVIDIA Jetson Xavier NX platform with the NN-Q40XT50 photoelectric turntable, achieving real-time detection and tracking of LSS unmanned aerial vehicles targets. The system exhibits reliable performance across different outdoor scenarios, achieving an average processing speed of 50 FPS and detection and tracking accuracy of 62% on average.
6.2 Future Work
Several research directions warrant further investigation in the future:
(1) Further improvement of small-target detection accuracy. Although the proposed algorithm achieves significant improvements, missed detections and false detections still occur under extreme long-distance or severe occlusion conditions. Future work could explore adaptive slicing-assisted hyper inference techniques and multi-dimensional attention mechanisms to further enhance robustness.
(2) Joint optimization of detection and tracking. The current S-Bot-SORT algorithm improves data association but remains influenced by detection errors. Future research could investigate end-to-end joint training frameworks that cooperatively optimize both the detector and tracker, incorporating confidence feedback or trajectory prediction correction mechanisms to reduce identity switch rates.
(3) Multi-target and similar-target interference handling. The current system primarily focuses on single-target scenarios. Future work could explore graph neural network-based global data association and multi-modal feature fusion for high-density, similar-target interference scenarios.
(4) Embedded system optimization and deployment expansion. While the current system achieves real-time performance on Jetson Xavier NX, there is still room for improvement in energy consumption and deployment scale. Future work could investigate network pruning, quantization acceleration, and edge-cloud collaborative inference to achieve efficient, low-power deployment.
(5) Long-term intelligent perception and autonomous response. From a broader perspective, the detection and tracking algorithms could be integrated with autonomous decision-making and intelligent response mechanisms to construct a comprehensive low-altitude UAV intelligent perception system, enabling autonomous perception, prediction, and response capabilities for drone safety regulation and urban low-altitude intelligent management.