In this work, we present CMambaSE_YOLOv11, a novel detection framework specifically designed for drone-based traffic object detection. Unmanned aerial vehicle (UAV) imagery introduces unique challenges, including small object sizes, severe occlusions, complex backgrounds, and large scale variations. To address these issues, we integrate state-space models with convolutional networks and multi-scale adaptive detection heads. Our approach significantly improves detection accuracy while maintaining real-time performance, making it suitable for intelligent transportation systems that rely on drone technology.
1. Introduction
Drone technology has become a cornerstone of modern urban traffic monitoring, offering flexibility, wide coverage, and cost efficiency. However, detecting traffic participants such as pedestrians, vehicles, and cyclists from aerial views remains difficult due to the small scale of objects, dense distribution, and varying illumination conditions. Traditional object detectors like YOLO series struggle in these scenarios because they lack effective long-range dependency modeling and adaptive multi-scale fusion.
Recent advances in state-space models, particularly Mamba, have demonstrated linear-time sequence modeling and strong global context capture. Meanwhile, channel attention mechanisms like squeeze-and-excitation (SE) can recalibrate feature responses adaptively. Combining these with convolutional backbones offers a promising direction for drone image analysis. In this paper, we propose three core contributions to enhance drone technology for traffic object detection:
- We design MambaSE, which integrates generalized Hilbert scanning with Mamba and SE block to jointly model spatial continuity and channel importance.
- We propose CMambaSE, a dynamic weighted fusion module that adaptively combines local convolutional features and global state-space features, replacing the conventional C3k2 module.
- We introduce an optimized FASFFHead detection head that employs spatial-aware fusion weights and small-object bias to improve multi-scale detection, especially for tiny objects.
Experimental results on two drone technology benchmarks, VisDrone2019 and UAVTrafficControlDataset, demonstrate that our model outperforms existing YOLO variants and state-of-the-art lightweight detectors.
2. Related Work
Drone technology has driven numerous improvements in object detection. Early works focused on enhancing CNN architectures with feature pyramids and attention mechanisms. For instance, FO-YOLO introduced a feature augmentation module for multi-scale context. Transformer-based methods like RT-DETR provided global receptive fields but suffered from high computational costs. Recently, state-space models have emerged as efficient alternatives. Mamba YOLO replaced self-attention with selective SSM, achieving competitive accuracy with lower latency. However, most existing Mamba-based detectors apply direct flattening of 2D features, losing spatial locality. Our work addresses this limitation by employing Hilbert scanning to preserve neighborhood structures, which is critical for detecting small objects in drone imagery.
3. Proposed Method
3.1 Overview of CMambaSE_YOLOv11
As shown in Figure 1, our network consists of a backbone, neck, and head. The backbone adopts a modified CSPDarknet structure where we replace all C3k2 modules with CMambaSE. The neck integrates multi-scale feature fusion enhanced by additional CMambaSE modules. The head uses the optimized FASFFHead for final predictions. This architecture fully leverages the complementary strengths of convolutions and state-space models tailored for drone technology.
3.2 MambaSE: Spatial-Channel Joint State-Space Modeling
Given an input feature map $F \in \mathbb{R}^{B \times C \times H \times W}$, we first apply a generalized Hilbert space-filling scan to convert it into a 1D sequence while preserving local spatial continuity. The scan function $G$ maps 2D coordinates to a 1D index $k$:
$$k = G(x, y \mid o, a, b)$$
where $o$ is the origin of the sub-region, $a$ and $b$ are the primary and orthogonal directions. Multi-direction scans (rotations and flips) are used for robustness. The sequence is then fed into the Mamba block, which models long-range dependencies via the state-space equations:
$$h_t = A_t h_{t-1} + B_t x_t$$
$$y_t = C_t h_t$$
After obtaining the output sequence, we apply an inverse Hilbert scan to restore the 2D grid structure. The restored feature map is then passed through an SE block to recalibrate channel responses:
$$z = \frac{1}{H \times W} \sum_{i=1}^{H}\sum_{j=1}^{W} F(i,j)$$
$$s = \sigma(W_2 \delta(W_1 z))$$
$$F’ = s \cdot F$$
This MambaSE module achieves both spatial context modeling and channel importance weighting, significantly enhancing small object features in drone technology.
3.3 CMambaSE: Dynamic Weighted Fusion of Convolution and MambaSE
The CMamba module (Figure 2) takes an input $X$ and processes it through two parallel branches:
- Local branch (C3k2): Extracts fine-grained local features.
- Global branch (MambaSE): Captures long-range context.
Let $F_{\text{local}} = \text{C3k2}(X)$ and $F_{\text{global}} = \text{MambaSE}(F_{\text{local}})$. A lightweight dynamic weight generator computes an adaptive fusion weight $\alpha$:
$$\alpha = \sigma(\text{MLP}(\text{GAP}(F_{\text{local}})))$$
where $\text{GAP}$ is global average pooling. The fused feature is then:
$$F_{\text{fusion}} = \alpha \cdot \gamma_{\text{c3k2}} \cdot F_{\text{local}} + (1-\alpha) \cdot \gamma_{\text{mamba}} \cdot F_{\text{global}}$$
Here $\gamma_{\text{c3k2}}$ and $\gamma_{\text{mamba}}$ are learnable scaling parameters to stabilize early training. A residual connection with DropPath enhances regularization. This dynamic fusion adapts to varying input complexity, effectively balancing local details and global semantics for drone technology applications.
3.4 Optimized FASFFHead Detection Head
Traditional detection heads fail to handle the extreme scale variation in drone imagery. We enhance the FASFF head with:
- Spatially-aware weight generation: For each FASFF level, we generate adaptive fusion weights using a lightweight depthwise separable convolution:
$$W_i = \text{DWConv}_{3\times3}(\text{Conv}_{1\times1}(X_i))$$
- Small-object bias: For high-resolution levels (level 2 and 3), we add a bias vector $B = [0.6, 0.3, 0.1]$ before softmax to increase the contribution of shallow features.
- Enhanced regression: The DFL regression dimension is increased from 16 to 20 for finer boundary prediction.
- Deeper classification branch: We use six convolutional layers (depthwise + standard) instead of two.
The fused feature at each level is computed as:
$$F_{\text{fused}} = \sum_{i=0}^{2} \hat{W}_i \cdot \text{Resize}(X_i)$$
where $\hat{W} = \text{softmax}(W + B)$.
This head significantly improves small object recall and localization accuracy in drone technology tasks.
4. Experiments
4.1 Experimental Setup
We conducted all experiments on an NVIDIA RTX4060 GPU with PyTorch and CUDA 11.8. Training hyperparameters are listed in Table 1.
| Parameter | Value |
|---|---|
| Image size | 640×640 |
| Batch size | 2 |
| Epochs | 150 |
| Optimizer | AdamW |
| Weight decay | 0.0005 |
| Cosine LR | True |
| Initial LR | 0.001 |
| Final LR | 0.0001 |
4.2 Datasets
We evaluate on two drone technology benchmarks:
- VisDrone2019: 10,209 images, 10 object classes, over 260k instances. Over 70% are small objects, making it a challenging testbed for drone traffic detection.
- UAVTrafficControlDataset: 3,717 images from three Lithuanian cities, annotated with 4 classes (Car, Truck, Pedestrian, Bus). It covers diverse traffic scenes and flight altitudes.
4.3 Ablation Studies
We conduct ablation on VisDrone2019 using YOLOv11n as baseline (Table 2).
| Baseline | CMambaSE | FASFFHead | Optimized FASFFHead | P | R | mAP50 | mAP50-95 |
|---|---|---|---|---|---|---|---|
| YOLOv11n | – | – | – | 0.492 | 0.189 | 0.343 | 0.206 |
| YOLOv11n | ✓ | – | – | 0.533 | 0.240 | 0.386 | 0.235 |
| YOLOv11n | – | ✓ | – | 0.523 | 0.227 | 0.375 | 0.228 |
| YOLOv11n | ✓ | ✓ | – | 0.537 | 0.268 | 0.401 | 0.243 |
| YOLOv11n | ✓ | – | ✓ | 0.550 | 0.256 | 0.403 | 0.248 |
Adding CMambaSE alone improves mAP50 from 0.343 to 0.386. Combining CMambaSE and optimized FASFFHead yields the best performance: mAP50=0.403, mAP50-95=0.248, and precision=0.55. This validates the effectiveness of each component in enhancing drone technology detection.
4.4 Comparison of Fusion Mechanisms
We compare dynamic weight fusion with feature-adaptive gating (Table 3). Despite slight recall improvement from gating, dynamic fusion achieves competitive accuracy with lower computational cost.
| Mechanism | P | R | mAP50 | mAP50-95 | Params(M) | GFLOPs | Latency(ms) | FPS |
|---|---|---|---|---|---|---|---|---|
| Dynamic Weight | 0.550 | 0.256 | 0.403 | 0.248 | 7.49 | 26.1 | 33.71 | 29.66 |
| Gating | 0.548 | 0.265 | 0.405 | 0.251 | 7.85 | 27.6 | 35.62 | 28.07 |
Dynamic weight fusion is chosen as default for its balance of performance and efficiency for drone technology applications.
4.5 Comparison of Scanning Strategies
Table 4 compares Hilbert scanning with row-major scanning under identical parameters. Hilbert scanning yields higher precision and mAP, though with slightly higher latency due to non-contiguous memory access.
| Scan Strategy | P | R | mAP50 | mAP50-95 | Params(M) | GFLOPs | Latency(ms) | FPS |
|---|---|---|---|---|---|---|---|---|
| Row-Major | 0.535 | 0.267 | 0.400 | 0.247 | 7.49 | 26.1 | 30.21 | 33.11 |
| Hilbert | 0.550 | 0.256 | 0.403 | 0.248 | 7.49 | 26.1 | 33.71 | 29.66 |
The Hilbert scan better preserves spatial topology, which is crucial for small object detection in drone technology.
4.6 Comparison with State-of-the-Art on VisDrone2019
Table 5 compares our model with leading detectors on VisDrone2019. Our approach achieves the highest mAP50 (0.403) and mAP50-95 (0.248) among lightweight models, demonstrating superior adaptation to drone technology challenges.
| Model | P | R | mAP50 | mAP50-95 | GFLOPs | Params |
|---|---|---|---|---|---|---|
| SSD | 0.211 | 0.358 | 0.240 | 0.119 | 63.2 | 12.30M |
| Faster-R-CNN | 0.346 | 0.368 | 0.309 | 0.131 | 370.0 | 63.20M |
| YOLOv5n | 0.505 | 0.180 | 0.343 | 0.206 | 7.2 | 2.5M |
| YOLOv8n | 0.523 | 0.194 | 0.359 | 0.215 | 8.2 | 3.0M |
| YOLOv10n | 0.537 | 0.150 | 0.345 | 0.212 | 8.4 | 2.7M |
| YOLOv11n | 0.492 | 0.189 | 0.343 | 0.206 | 6.4 | 2.5M |
| YOLOv5_mamba | 0.517 | 0.384 | 0.401 | 0.237 | 50.9 | 12.9M |
| Ours | 0.550 | 0.256 | 0.403 | 0.248 | 26.1 | 7.49M |
Our model strikes an optimal balance between precision and recall, yielding the highest mAP despite lower recall than some Mamba-based models. This trade-off is beneficial for reliable detection in drone technology systems where false positives are costly.
4.7 Comparison on UAVTrafficControlDataset
To validate generalization, we evaluate on another drone imagery dataset (Table 6). Our model achieves the best mAP50 (0.909) and mAP50-95 (0.685), with the highest recall (0.892).
| Model | P | R | mAP50 | mAP50-95 |
|---|---|---|---|---|
| YOLOv5n | 0.894 | 0.794 | 0.821 | 0.614 |
| YOLOv8n | 0.933 | 0.808 | 0.878 | 0.668 |
| YOLOv11n | 0.891 | 0.768 | 0.810 | 0.612 |
| YOLOv5_mamba | 0.870 | 0.818 | 0.833 | 0.624 |
| Ours | 0.886 | 0.892 | 0.909 | 0.685 |
Our method demonstrates strong generalization across different drone technology datasets, confirming its robustness for real-world traffic monitoring.
4.8 Visualization Analysis
We provide qualitative comparisons on five typical drone scenes: multi-scale, vertical view, dense small objects, illumination variation, and nighttime. Our model consistently reduces missed detections and false positives, especially for small pedestrians and distant vehicles. Heatmap visualizations further show that our model activates more precisely on object regions while suppressing background noise, indicating better feature discriminability for drone technology tasks.
5. Conclusion
In this paper, we presented CMambaSE_YOLOv11, a novel detection framework tailored for drone-based traffic object detection. By integrating Hilbert scanning, state-space modeling, channel attention, dynamic feature fusion, and an optimized multi-scale detection head, our model effectively addresses the challenges of small object scale, occlusion, and complex backgrounds inherent in drone technology. Extensive experiments on two public benchmarks demonstrate consistent improvements over state-of-the-art lightweight detectors, achieving up to 17.5% mAP50 improvement on VisDrone2019. Future work will explore model compression techniques such as pruning and knowledge distillation to further enhance inference speed while maintaining accuracy, enabling real-time deployment on edge devices in drone technology systems.