In recent years, the rapid proliferation of low-altitude unmanned aerial vehicles across both military and civilian domains has introduced increasingly prominent security hazards. The unlawful intrusion of unmanned aerial vehicles into controlled airspace has been frequently reported, posing serious threats to public safety and national security. Consequently, the realization of efficient identification and continuous tracking of unmanned aerial vehicles has become a fundamental prerequisite for implementing diverse countermeasure operations. In this thesis, I conduct an in-depth investigation into the key technologies of detection and tracking for unmanned aerial vehicles, with the objective of addressing the challenges of small target size, severe scale variation, complex background interference, and temporary target disappearance that are commonly encountered in real-world scenarios.
1. Introduction and Research Background
With the advancement of aviation technology and the gradual relaxation of airspace management policies, unmanned aerial vehicles have achieved large-scale application in numerous industries, owing to their low cost, high maneuverability, and excellent imaging capabilities. In military applications, unmanned aerial vehicles have become critical combat assets capable of performing reconnaissance, electronic interference, precision strikes, and communication relay missions. In industrial production, they are widely deployed for logistics transportation, surveying and mapping, agricultural plant protection, inspection, and emergency rescue. In the consumer market, they have gained popularity for recreational flight and aerial photography. According to recent market research, the global anti-drone market is projected to reach USD 7.05 billion by 2029, with a compound annual growth rate of 26.7% from 2024.
However, the widespread utilization of unmanned aerial vehicles has concurrently escalated the risk of malicious use, including privacy infringement, terrorist attacks, and threats to critical infrastructure. Traditional radar-based detection systems often suffer from degraded performance when detecting small unmanned aerial vehicles, which typically have small radar cross-sections, are constructed from composite materials, and fly at relatively low altitudes. RF-based and acoustic sensing methods also exhibit limitations in complex electromagnetic environments. In this context, vision-based detection methods leveraging deep learning have demonstrated exceptional potential, offering superior capability in extracting precise positional, dimensional, and appearance information of aerial targets. The development of robust visual detection and tracking algorithms for unmanned aerial vehicles has thus become a research hotspot of significant academic and practical value.
The primary technical challenges addressed in this research are as follows. First, when detecting distant unmanned aerial vehicles, the target occupies an extremely small proportion of the image, and its appearance features are often indistinct, resulting in poor detection performance of generic detection models. Second, the motion of unmanned aerial vehicles can be highly dynamic, and the presence of similar background objects can cause tracking drift or loss. Third, the deployment of deep learning models on resource-constrained edge devices imposes stringent requirements on model lightweighting and computational efficiency.
2. Related Work
2.1 Deep Learning Foundations
Deep learning, as a critical branch of machine learning, leverages multi-layer neural network architectures to directly learn feature representations from data. The fundamental building block of deep neural networks is the artificial neuron, which computes a weighted sum of inputs followed by a nonlinear activation transformation. The output of a single perceptron can be expressed as:
$$y = f\left(\sum_{i=1}^{n} w_i x_i + b\right)$$
where \(x_i\) represents the input feature vector, \(w_i\) denotes the corresponding weight parameter, \(b\) is the bias, and \(f(\cdot)\) is the activation function. Through iterative optimization using backpropagation, the network parameters are updated to minimize the discrepancy between predicted outputs and ground-truth labels, thereby enabling the model to learn complex mappings from input data to desired outputs.
Convolutional neural networks constitute a specialized class of deep neural networks that have achieved remarkable success in image processing tasks. A typical CNN architecture consists of convolutional layers, pooling layers, and fully connected layers. The convolution operation can be formulated as:
$$Y = \sigma(X \ast W + b)$$
where \(X\) is the input feature map, \(W\) represents the convolution kernel weights, \(b\) is the bias term, and \(\sigma\) denotes the nonlinear activation function. The spatial dimension of the output feature map is determined by the input size \(I\), kernel size \(K\), padding \(P\), and stride \(S\):
$$O = \frac{I – K + 2P}{S} + 1$$
Pooling layers serve to downsample feature maps, reducing spatial resolution while preserving important features. The two principal pooling operations are max pooling and average pooling, which extract the maximum response and the mean value within each local region, respectively.
2.2 Object Detection Methods
Object detection requires both classification and localization of objects within an image. Traditional detection approaches relied on hand-crafted features such as SIFT, HOG, and Haar-like features, together with classifiers like SVM and Adaboost. Although these methods achieved reasonable performance under constrained conditions, they were generally limited in handling complex scenes and significant appearance variations.
The advent of deep learning has revolutionized the field of object detection. Two-stage detectors, represented by the R-CNN family, first generate candidate regions and subsequently perform classification and bounding-box regression on these regions. Faster R-CNN introduced the Region Proposal Network, which dramatically improved detection speed while maintaining high accuracy. Single-stage detectors, exemplified by the YOLO series and SSD, directly predict object categories and locations from the input image in a single forward pass, achieving real-time inference speeds. Table 1 summarizes the key characteristics of the YOLO series evolution.
| Year | Version | Key Characteristics |
|---|---|---|
| 2016 | YOLOv1 | Unified real-time detection framework; limitations in small-object detection |
| 2017 | YOLOv2 | Anchor mechanism, batch normalization, and multi-scale training |
| 2018 | YOLOv3 | Darknet-53 backbone, multi-scale prediction, logistic regression |
| 2020 | YOLOv4 | CSPDarknet53, PANet, and Mosaic data augmentation |
| 2020 | YOLOv5 | PyTorch-based implementation, flexible model scaling |
| 2023 | YOLOv8 | Anchor-free detection, C2f module, decoupled head, TAL strategy |
The YOLOv8 architecture employs a CSPDarknet-like backbone with the C2f module replacing the C3 module to improve gradient flow and feature reuse. The neck utilizes a PAN-FPN structure for multi-scale feature fusion, while the head adopts a decoupled design separating classification and regression tasks. The loss function combines Varifocal Loss for classification and CIoU Loss with Distribution Focal Loss for regression.
2.3 Target Tracking Methods
Visual target tracking aims to continuously locate a specific target across video frames. Generative methods, such as particle filters and mean shift, construct appearance models to represent the target. Discriminative methods train binary classifiers to separate the target from the background. The Kalman filter and Hungarian algorithm are two fundamental components in modern tracking frameworks.
The Kalman filter provides an optimal recursive state estimation for linear dynamic systems corrupted by Gaussian noise. The state prediction and update equations are given by:
$$X_{k+1} = A X_k + W_k$$
$$Z_k = H X_k + V_k$$
where \(A\) denotes the state transition matrix, \(H\) is the observation matrix, and \(W_k\), \(V_k\) represent process and observation noise with covariance matrices \(Q\) and \(R\), respectively. The prediction of the state error covariance is:
$$P_{k+1|k} = A P_{k|k} A^{\mathrm{T}} + Q$$
With the introduction of the Kalman gain \(K_{k}\), the optimal state estimate is obtained as:
$$X_{k+1|k+1} = X_{k+1|k} + K_{k+1}(Z_{k+1} – H X_{k+1|k})$$
$$K_{k+1} = P_{k+1|k} H^{\mathrm{T}} (H P_{k+1|k} H^{\mathrm{T}} + R)^{-1}$$
The Hungarian algorithm solves the assignment problem in polynomial time and is widely employed to associate detections with trajectories across frames. It constructs a cost matrix representing the similarity between predicted and detected bounding boxes, and subsequently computes the optimal one-to-one matching.
3. PCCS-YOLOv8 for Unmanned Aerial Vehicles Detection
3.1 Overview of the Proposed Method
To address the challenges of small-object detection for unmanned aerial vehicles in complex backgrounds, I propose the PCCS-YOLOv8 algorithm. The overall architecture, illustrated conceptually in Figure 1, enhances the baseline YOLOv8 model through several key innovations: the addition of a P2 small-object detection layer, the integration of a Spatial Pyramid Pooling module with Efficient Layer Aggregation Networks, the incorporation of the Convolutional Block Attention Module, and the construction of a lightweight CSPPC module employing partial convolutions.
3.2 P2 Small-Object Detection Layer
The baseline YOLOv8 model employs multi-scale detection heads with downsampling factors of 8x, 16x, and 32x. The minimum detection feature map has a resolution of 8×8 pixels, which limits the detection capability for objects smaller than this threshold. To overcome this limitation, I introduce the P2 detection head, which utilizes a 4x downsampled feature map. This enhancement enables the model to detect objects with as few as 4×4 pixels, significantly improving the sensitivity to small targets. Table 2 shows the detection resolution ranges for different detection heads at an input image size of 640×640 pixels.
| Detection Head | Feature Map Size | Minimum Object Size |
|---|---|---|
| P2 (newly added) | 160×160 | ≥4×4 pixels |
| P3 | 80×80 | ≥8×8 pixels |
| P4 | 40×40 | ≥16×16 pixels |
| P5 | 20×20 | ≥32×32 pixels |
The introduction of the P2 detection layer establishes a feature pyramid with four detection scales, enabling the network to effectively handle targets of varying sizes. The feature fusion is accomplished through a bottom-up pathway and a top-down pathway that integrate contextual information from different levels.
3.3 C2f Module Improvment with Partial Convolution
The addition of the P2 detection layer inevitably increases model complexity and computational costs. To balance detection performance with efficiency, I propose the CSPPC (Cross Stage Partial with Partial Convolution) module as a replacement for the original C2f module in the neck network. The core innovation lies in utilizing Partial Convolution to reduce redundant computations.
Partial Convolution performs convolution operations on only a subset of input channels while preserving the remaining channels through identity mapping. Given an input feature map \(X \in \mathbb{R}^{C \times H \times W}\), the convolution is applied to a subset of channels \(X_{p} \in \mathbb{R}^{c_p \times H \times W}\) with a ratio \(r = c_p / C\). The output can be expressed as:
$$Y = [X_{p} \ast W_{p}; \quad X_{u}]$$
where \(W_{p}\) is the convolution kernel applied to the selected channels, and \(X_{u}\) represents the unprocessed channels passing through identity mapping. This approach achieves a significant reduction in FLOPs. The computational complexity comparison is presented in Table 3, where standard convolution requires \(k^2 C^2 H W\) operations, while Partial Convolution requires only \(k^2 c_p^2 H W\), with notable savings when \(r < 1\).
| Convolution Type | FLOPs | Memory Access |
|---|---|---|
| Standard Convolution | \(k^2 C^2 H W\) | \(C H W\) |
| Depthwise Convolution | \(k^2 C H W\) | \(C H W\) |
| Partial Convolution | \(k^2 c_p^2 H W\) | \((c_p + C) H W\) |
In the proposed CSPPC module, the Bottleneck layers in the original C2f structure are replaced with Bottleneck variants using Partial Convolution. The module follows a dual-pathway architecture that splits the input, processes features through multiple Bottleneck layers, and concatenates the outputs. This design enhances feature representation capabilities while substantially reducing the number of parameters and floating-point operations required.
3.4 CBAM Attention Mechanism
To enable the network to focus on important features of small unmanned aerial vehicles, I introduce the Convolutional Block Attention Module. CBAM sequentially applies channel and spatial attention to refine feature representations. Given an input feature map \(F\), the channel attention is computed by:
$$M_{c}(F) = \sigma\left(\mathrm{MLP}(\mathrm{AvgPool}(F)) + \mathrm{MLP}(\mathrm{MaxPool}(F))\right)$$
where \(\sigma\) denotes the sigmoid function, and the shared MLP has a hidden layer of size \(C/r\). The channel-refined feature is then passed to the spatial attention module:
$$M_{s}(F’) = \sigma\left(f^{7 \times 7}([\mathrm{AvgPool}(F’); \mathrm{MaxPool}(F’)])\right)$$
The overall CBAM operation can be summarized as:
$$F’ = M_{c}(F) \otimes F$$
$$F” = M_{s}(F’) \otimes F’$$
In my implementation, CBAM modules are strategically inserted into the backbone, specifically after deep C2f modules, and in the neck network at cross-scale connection nodes. This placement enables the model to suppress background interference and amplify salient target information, thereby improving detection accuracy for small unmanned aerial vehicles in cluttered scenes.
3.5 SPPELAN Module
The SPPF module in the original YOLOv8 facilitates multi-scale feature extraction through cascaded max-pooling operations. Considering the characteristics of small object detection, I propose the SPPELAN module, which combines the strengths of spatial pyramid pooling and efficient layer aggregation networks. The structure processes features through parallel max-pooling operations of varying kernel sizes, followed by a transition that merges multi-scale information. The network benefits from hierarchical feature aggregation while maintaining computational efficiency. The pooling operation is formulated as:
$$y_{i,j} = \max_{m,n} (x_{i \cdot s + m, \; j \cdot s + n})$$
The integration of SPPELAN in the final stage of the backbone enhances the model’s receptive field and facilitates better multi-scale feature extraction for both small and large unmanned aerial vehicles in changing environmental conditions.
3.6 Experiments and Results for Detection
The experiments were conducted utilizing a dataset integrating TIB-UAV, Anti-UAV, and self-collected unmanned aerial vehicles samples. The dataset comprises 7,785 images with diverse scenes including sky, cloud, and high-rise building backgrounds. The dataset statistics indicate that approximately 91% of the annotated targets are small or medium-sized, reflecting the challenges inherent in distant unmanned aerial vehicles detection. All experiments were performed on an NVIDIA GeForce RTX 4060 GPU with 8GB memory, using the PyTorch framework.
The evaluation metrics employed include mean Average Precision at IoU threshold 0.5 (mAP@0.5), mAP across IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95), parameters (Params), floating-point operations (FLOPs), and frames per second (FPS). The performance comparison of various detection algorithms is presented in Table 4.
| Algorithm | mAP@0.5 (%) | mAP@0.5:0.95 (%) | FLOPs (G) | Params (M) |
|---|---|---|---|---|
| YOLOv5 | 89.3 | 45.6 | 7.2 | 2.51 |
| YOLOv8 | 91.1 | 48.4 | 8.2 | 3.01 |
| YOLOv10n | 90.2 | 49.1 | 8.4 | 2.71 |
| RT-DETR-L | 92.2 | 46.4 | 103.4 | 32 |
| PCCS-YOLOv8 | 94 | 50.6 | 11.1 | 2.81 |
The experimental results demonstrate that the proposed PCCS-YOLOv8 algorithm achieves superior performance across all evaluation metrics. Compared with the baseline YOLOv8, the mAP@0.5 improves by 2.9 percentage points, reaching 94%, while mAP@0.5:0.95 improves by 2.2 percentage points, reaching 50.6%. Although the computational cost increases due to the addition of the P2 detection layer, the lightweight CSPPC module effectively mitigates this overhead. The parameter count is reduced to 2.81M, which is lower than the baseline YOLOv8’s 3.01M.
I further conducted ablation studies to analyze the contribution of each improvement component. The results are summarized in Table 5.
| P2 | CSPPC | CBAM | SPPELAN | mAP@0.5 (%) | FLOPs (G) | Params (M) |
|---|---|---|---|---|---|---|
| – | – | – | – | 91.1 | 8.2 | 3.01 |
| ✓ | – | – | – | 94 | 12.2 | 2.92 |
| – | ✓ | – | – | 91.4 | 7.2 | 2.57 |
| – | – | ✓ | – | 91.9 | 8.2 | 3.09 |
| – | – | – | ✓ | 91.1 | 8.1 | 2.85 |
| – | – | ✓ | ✓ | 91.7 | 8.2 | 3.02 |
| ✓ | – | ✓ | ✓ | 94.2 | 12.5 | 3.26 |
| ✓ | ✓ | ✓ | ✓ | 94 | 11.1 | 2.81 |
The ablation results reveal that the P2 detection layer provides the most significant performance gain, improving mAP@0.5 by 2.9 percentage points. The CSPPC lightweight module contributes to a reduction of 11.2% in FLOPs and 13.8% in paraters while maintaining comparable detection accuracy. The CBAM attention mechanism further improves both mAP@0.5 and mAP@0.5:0.95, confirming its effectiveness in focusing on small-object features. The complete PCCS-YOLOv8 model achieves a good balance between detection accuracy and computational efficiency.
A comparative analysis of attention mechanisms is presented in Table 6, illustrating that CBAM outperforms other attention modules including ECA, GAM, and SE in terms of detection accuracy.
| Attention Module | mAP@0.5 (%) | mAP@0.5:0.95 (%) | FLOPs (G) | Params (M) |
|---|---|---|---|---|
| None | 91.1 | 48.4 | 8.2 | 3.01 |
| CBAM | 91.9 | 48.7 | 8.2 | 3.09 |
| ECA | 91.5 | 48.9 | 8.2 | 3.10 |
| GAM | 90 | 47.7 | 18.2 | 11.7 |
| SE | 91.6 | 48.3 | 8.1 | 3.01 |
Qualitative detection results on challenging small-target scenes confirmed that the proposed PCCS-YOLOv8 algorithm successfully detects unmanned aerial vehicles that the baseline YOLOv8 misses, and achieves higher detection confidence scores when both algorithms successfully identify the target. The algorithm also demonstrates robust detection performance in complex scenarios where the target shares color similarity with the background.
4. PK-ByteTrack for Unmanned Aerial Vehicles Tracking
4.1 ByteTrack Framework
ByteTrack is a popular tracking-by-detection algorithm that introduces the BYTE data association strategy. Unlike traditional methods that discard low-confidence detection boxes, ByteTrack effectively exploits both high-confidence and low-confidence detections to improve tracking robustness. The algorithm maintains track states and employs Kalman filtering for motion prediction. Each trajectory is associated with a temporal visibility counter that increments during prediction and resets upon successful matching.
The data association in ByteTrack involves a two-stage matching process. In the first stage, high-confidence detections are matched with existing tracked objects. In the second stage, low-confidence detections are associated with unmatched tracks, which helps handle occlusion and motion blur. The Hungarian algorithm is used to compute optimal assignments based on IoU similarity between predicted boxes and detection boxes.
4.2 Improved Kalman Filter for UAV Tracking
In the original ByteTrack, the state vector space is \(\mathbb{R}^{8}\) with the form:
$$X = [u, v, r, h, \dot{u}, \dot{v}, \dot{r}, \dot{h}]^{\mathrm{T}}$$
where \(u\) and \(v\) denote the target bounding box center coordinates, \(r\) represents the aspect ratio, \(h\) denotes the height, and the dotted variables represent their corresponding velocities. In the context of small unmanned aerial vehicles in infrared imagery, the aspect ratio assumption can degrade performance when target appearance changes rapidly.
To address this limitation, I propose an enhanced state representation that directly predicts the bounding box width and height independently:
$$X = [u, v, w, h, \dot{u}, \dot{v}, \dot{w}, \dot{h}]^{\mathrm{T}}$$
This modification allows the Kalman filter to independently model variations in both width and height, which is particularly beneficial for unmanned aerial vehicles performing agile maneuvers where the bounding box dimensions fluctuate considerably. The noise covariance matrices are optimized as follows:
$$Q = \mathrm{diag}(\sigma_p^2 w^2, \sigma_p^2 h^2, \sigma_p^2 w^2, \sigma_p^2 h^2, \sigma_v^2 w^2, \sigma_v^2 h^2, \sigma_v^2 w^2, \sigma_v^2 h^2)$$
where \(\sigma_p\) and \(\sigma_v\) are position and velocity noise factors, and the dimensions are adaptively scaled with the target width \(w\) and height \(h\). This provides more accurate uncertainty modeling for targets of varying sizes.
4.3 PK-ByteTrack Framework
I propose the PK-ByteTrack algorithm by integrating the PCCS-YOLOv8 detector with an improved Kalman filter within the ByteTrack framework. The overall framework can be described through the following main components: the PCCS-YOLOv8 detector provides high-quality detection results, the enhanced Kalman filter performs trajectory prediction, and the Hungarian algorithm accomplishes optimal matching between predicted and detected boxes. The detection boxes are classified into high-confidence and low-confidence groups during preprocessing.
The tracking process follows a structured pipeline. First, the provided video frame is fed into the PCCS-YOLOv8 detector, which outputs object boxes, confidence scores, and class labels. The detection boxes are partitioned based on confidence thresholds \(\tau_{high}\) and \(\tau_{low}\). During the primary matching stage, high-confidence detections are matched against existing tracks. In the secondary matching stage, low-confidence detections are associated with unmatched tracks. Successful matches result in track updates, while unmatched high-confidence detections initialize new tracks. Tracks that remain unmatched for a certain threshold duration are terminated.
4.4 Tracking Experiments and Results
The tracking experiments were conducted on a self-constructed MOT-format infrared unmanned aerial vehicles dataset containing four video sequences, each comprising 1,000 frames. These sequences cover challenging scenarios including cirrus cloud backgrounds, building interference, target rotation, and rapid descent. The Multiple Object Tracking Accuracy metric is formulated as:
$$MOTA = 1 – \frac{FN + FP + IDSW}{GT}$$
where \(FN\), \(FP\), and \(IDSW\) represent the number of false negatives, false positives, and identity switches, respectively, and \(GT\) is the total number of ground-truth objects.
Table 7 presents the tracking results comparing the proposed PK-ByteTrack with DeepSORT, BotSORT, and ByteTrack.
| Sequence | Algorithm | FP ↓ | FN ↓ | IDSW ↓ | MOTA ↑ | FPS |
|---|---|---|---|---|---|---|
| Sequence 1 | DeepSORT | 15 | 124 | 103 | 75.8 | 31.88 |
| BotSORT | 9 | 90 | 127 | 77.4 | 50.33 | |
| ByteTrack | 5 | 87 | 128 | 78 | 52.58 | |
| PK-ByteTrack | 0 | 83 | 115 | 80.2 | 49.6 | |
| Sequence 2 | DeepSORT | 33 | 62 | 51 | 85.4 | 29.3 |
| BotSORT | 1 | 64 | 102 | 83.3 | 52.91 | |
| ByteTrack | 3 | 60 | 89 | 84.8 | 50.94 | |
| PK-ByteTrack | 3 | 46 | 46 | 90.5 | 48.22 | |
| Sequence 3 | DeepSORT | 34 | 119 | 126 | 72.1 | 29.44 |
| BotSORT | 5 | 126 | 137 | 73.2 | 51.68 | |
| ByteTrack | 6 | 102 | 149 | 74.3 | 46.9 | |
| PK-ByteTrack | 4 | 89 | 143 | 76.4 | 37.88 | |
| Sequence 4 | DeepSORT | 63 | 103 | 78 | 75.9 | 31.73 |
| BotSORT | 36 | 114 | 116 | 73.4 | 59.02 | |
| ByteTrack | 23 | 110 | 83 | 78.4 | 53.34 | |
| PK-ByteTrack | 18 | 87 | 60 | 83.5 | 52.71 |
The experimental results demonstrate that the PK-ByteTrack algorithm achieves consistent improvements in MOTA across all four sequences compared with the baseline ByteTrack. The improvements are 2.2%, 5.7%, 2.1%, and 5.1% for sequences 1 through 4, respectively. Notably, the algorithm exhibits significant reductions in false negatives and identity switches, particularly in Sequence 2 and Sequence 4, which validates the effectiveness of the improved Kalman filter in maintaining trajectory continuity during complex maneuvers.
Qualitative analysis of the tracking results further confirmed the advantages of the proposed approach. In a scenario where the target unmanned aerial vehicle shared a similar color distribution with the building background, the baseline ByteTrack experienced target loss in frames 110-111 due to decreased detection confidence. In contrast, the PK-ByteTrack successfully maintained tracking continuity throughout the sequence, demonstrating enhanced robustness against background interference. The ID-switch analysis on cluttered cloud backgrounds indicated that the improved algorithm significantly reduced frequent identity changes, yielding more stable and reliable tracking outputs.
5. Embedded Platform Implementation
5.1 Hardware Selection
To validate the practical applicability of the proposed algorithms, I implemented the optimized detection and tracking system on an embedded platform. Table 8 presents a comparison of several candidate embedded platforms.
| Parameter | Orange Pi 5 Pro | Raspberry Pi 5 | Jetson Orin Nano |
|---|---|---|---|
| CPU | 4×A76 + 4×A55 | 4×A76 | 6×A78AE |
| GPU | ARM Mali-G610 MP4 | VideoCore VII | Ampere GPU |
| Memory | LPDDR5 up to 16GB | LPDDR4X 4/8GB | LPDDR5 4/8GB |
| NPU | 6 TOPS | External | 40 TOPS |
| Network | 2.5G Ethernet | 1G Ethernet | 1G Ethernet |
| Price | 700-1000 RMB | 600-800 RMB | 2000+ RMB |
| Power | 10-15W | 5-10W | 10-25W |
After careful consideration of computational capability, power consumption, cost, and expandability, I selected the Orange Pi 5 Pro featuring the Rockchip RK3588S octa-core processor. The platform provides 6 TOPS of NPU computing power, LPDDR5 memory support up to 16GB, and rich I/O interfaces, while maintaining a compact form factor suitable for field deployment.
For image acquisition, I adopted a camera based on the OV2710 sensor, which provides 2 million effective pixels, 1080p full high-definition resolution output, a 120-degree wide-angle field of view, and USB 2.0 plug-and-play connectivity. The camera parameters are summarized in Table 9.
| Attribute | Parameter |
|---|---|
| Sensor | OV2710 |
| Pixels | 2 million |
| Interface | USB 2.0 plug-and-play |
| Resolution | 1080p |
| Field of View | 120 degrees |
| CMOS Size | 1/2.7 inch |
The experimental deployment environment was established on a rooftop location at approximately 120 meters from the flight area, providing a realistic setting for outdoor unmanned aerial vehicles monitoring.
5.2 Model Conversion for Edge Deployment
The deployment of deep learning models on the Orange Pi 5 Pro requires the conversion of the trained PyTorch model into the RKNN format suitable for the Rockchip NPU. The conversion process comprises two stages.
In the first stage, the PyTorch model is converted to the ONNX format. This process involves defining input and output data types, associating the model graph with metadata, importing the operator set, and generating a serialized graph structure. The conversion can be represented as:
$$\mathcal{G}_{\mathrm{ONNX}} = \mathcal{T}(\mathcal{G}_{\mathrm{PyTorch}}, \mathcal{M})$$
where \(\mathcal{G}_{\mathrm{PyTorch}}\) represents the PyTorch computation graph and \(\mathcal{M}\) denotes the model metadata.
In the second stage, the ONNX model is compiled into the RKNN format. This involves constructing an RKNN object, configuring quantization parameters, loading the ONNX model, building the computational graph, and exporting the final RKNN file. The RKNN model format is specifically optimized for Rockchip NPU execution, providing significant inference acceleration compared with CPU-only execution.
5.3 System Development and Experimental Validation
The developed detection and tracking system employs the PyQT5 framework for the graphical user interface. The system consists of a model configuration module, an image detection module, and a video tracking module. The model configuration interface enables users to select different detection models, load custom weight files and class label files, and configure important parameters such as input image size, confidence threshold, and IoU threshold.
The image detection functionality allows users to upload images for offline analysis. When the system identifies an unmanned aerial vehicle, the detection results, including bounding boxes and confidence scores, are displayed. The video file tracking functionality supports the processing of recorded video sequences, with real-time display of detected target counts and tracking trajectories.
The real-time monitoring functionality connects the system to the OV2710 camera for live unmanned aerial vehicles tracking. When an unmanned aerial vehicles appears in the monitoring field of view, the system rapidly identifies and continuously tracks the target, as demonstrated in outdoor experiments where the target was positioned approximately 120 meters from the platform.
The deployment results confirmed that the proposed algorithm maintains efficient and stable target tracking performance in resource-constrained environments. The system successfully detects and tracks unmanned aerial vehicles in real environments, and the mean detection precision satisfies practical application requirements. This validation demonstrates the feasibility and scalability of the entire solution, opening a feasible approach for applying the proposed algorithms in actual industrial environments.
6. Conclusion and Future Work
This thesis has systematically investigated the detection and tracking techniques for unmanned aerial vehicles. The main contributions of this research are summarized as follows.
First, I proposed the PCCS-YOLOv8 algorithm for unmanned aerial vehicles detection. By introducing the P2 small-object detection layer, integrating the SPPELAN module, incorporating the CBAM attention mechanism, and constructing the lightweight CSPPC module, the proposed method significantly enhanced the detection accuracy for small unmanned aerial vehicles in complex environments. The experimental results demonstrated that mAP@0.5 reached 94% and mAP@0.5:0.95 reached 50.6%, representing improvements of 2.9% and 2.2% over the baseline YOLOv8, respectively, while maintaining a modest computational cost of 11.1 G FLOPs and 2.81 M parameters.
Second, I proposed the PK-ByteTrack algorithm for infrared unmanned aerial vehicles tracking. By integrating the PCCS-YOLOv8 detector with an improved Kalman filter that independently models target width and height, the proposed method achieved significant tracking performance improvements. The MOTA improved by 2.2%, 5.7%, 2.1%, and 5.1% across four challenging video sequences, demonstrating enhanced robustness against complex backgrounds and improved tracking stability.
Third, I implemented a complete unmanned aerial vehicles detection and tracking system on the embedded Orange Pi 5 Pro platform. The practical deployment validated the real-world applicability of the proposed algorithms, proving that efficient and stable target tracking can be achieved under resource-constrained conditions.
Regarding future research directions, several aspects warrant further investigation. Multi-modal information fusion, combining visible light, infrared, radar, and acoustic sensing technologies, can overcome the limitations of single-sensor approaches and enhance the robustness and adaptability of the detection system in complex operational environments. For edge computing optimization, dedicated neural network accelerators for unmanned aerial vehicles detection tasks and model compression techniques, including network pruning and knowledge distillation, can further reduce computational complexity and enable deployment in a broader range of embedded applications. Addressing these directions will contribute to the development of more effective and practical counter-drone systems, thereby enhancing the safety and security of low-altitude airspace.