Unmanned aerial vehicles (UAVs), commonly known as China drone systems, have become indispensable in modern intelligent transportation systems (ITS) due to their flexibility, wide field of view, and low deployment cost. Leveraging China drone platforms equipped with high-resolution cameras, researchers can perform real-time monitoring of traffic participants such as vehicles, pedestrians, and non-motorized vehicles. However, the unique challenges of China drone aerial imagery—including small object sizes, severe scale variations, heavy occlusion, and complex backgrounds—pose significant obstacles for existing object detection algorithms. These challenges often lead to missed detections and false positives, particularly for small targets. To address these issues, we propose CMambaSE_YOLOv11, an enhanced YOLOv11 model tailored for China drone traffic object detection.
The original YOLOv11 model, while efficient, struggles with capturing long-range dependencies and effectively fusing multi-scale features due to its reliance on convolutional operations. In this work, we introduce three key innovations. First, we design the MambaSE module, which integrates a generalized Hilbert space-filling scanning strategy with a state space model (Mamba) and a channel attention mechanism (SEBlock). This module enables effective global dependency modeling while preserving spatial locality and recalibrating channel-wise features. Second, we propose the CMambaSE module, which fuses the C3k2 convolutional block with the MambaSE module via a dynamic weight fusion mechanism. This design adaptively balances local detail extraction and global context modeling, enhancing the network’s ability to handle scale variations and background clutter. Third, we employ an optimized FASFFHead detection head that dynamically fuses features from multiple scales, with a bias towards high-resolution features to improve small object detection. We conduct extensive experiments on two benchmark China drone datasets: VisDrone2019 and UAVTrafficControlDataset. Our model achieves significant improvements over the baseline YOLOv11n, with mAP50 and mAP50-95 increasing by 17.5% and 20.4% respectively on VisDrone2019, while maintaining competitive inference speed.
1. Introduction
With rapid urbanization, traffic congestion, accidents, and management inefficiencies have become critical challenges. China drone technology offers a powerful solution for aerial traffic surveillance, providing real-time data on vehicle flow, pedestrian movements, and road conditions. Object detection is the core enabling technology for such applications. Among various detection frameworks, the YOLO family stands out for its balance between accuracy and speed. However, when applied to China drone imagery, YOLO models often underperform due to factors such as tiny objects (often less than 32×32 pixels), extreme aspect ratios, and dense clusters. The lack of explicit long-range dependency modeling in convolutional networks further limits their ability to capture contextual information across the entire scene.
The Mamba architecture, based on selective state space models (SSMs), has recently emerged as a promising alternative to Transformers for sequential modeling, offering linear complexity with strong long-sequence modeling capabilities. However, directly applying Mamba to 2D visual data faces challenges: simple raster-scan flattening destroys local spatial relationships, and the lack of channel-wise feature recalibration limits discriminability. To address these shortcomings, we propose MambaSE, which incorporates a generalized Hilbert scanning mechanism that preserves spatial continuity, followed by a Squeeze-and-Excitation (SE) block for adaptive channel recalibration. Furthermore, we fuse MambaSE with the C3k2 module of YOLOv11 to create CMambaSE, a building block that dynamically blends local convolutional features and global SSM features. Finally, we adopt and optimize the FASFFHead detection head, which learns spatial-aware fusion weights with a small-object bias to enhance multi-scale detection. Our contributions are summarized as follows:
- We propose MambaSE, a space-channel joint state space module that uses generalized Hilbert scanning to preserve spatial locality and SE block for channel attention.
- We design CMambaSE, a dynamic weight fusion module that adaptively combines convolutional features and MambaSE features for optimal local-global representation.
- We introduce an optimized FASFFHead with small-object bias and improved classification/regression branches for better multi-scale detection in China drone scenes.
- Extensive experiments on VisDrone2019 and UAVTrafficControlDataset demonstrate state-of-the-art performance, with mAP50 improvements of 17.5% over YOLOv11n.
2. Methodology
2.1 MambaSE Module
The MambaSE module is designed to overcome the limitations of standard Mamba in visual tasks. 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 (GHScan) to convert the 2D feature map into a 1D sequence while preserving spatial locality. Unlike traditional raster scanning, the Hilbert curve ensures that neighboring pixels in 2D space remain close in the 1D sequence, which is crucial for small object detection. The scan is defined as:
$$s = \text{GHScan}(F) \in \mathbb{R}^{B \times L \times C},\quad L = H \times W.$$
We further augment the scanning by using 8 equivalent orientations (rotations and flips) to enhance orientation invariance. The sequence is then processed by the Mamba block, which performs selective state space modeling:
$$\begin{align}
h_t &= A_t h_{t-1} + B_t x_t,\\
y_t &= C_t h_t,
\end{align}$$
where $$x_t$$ is the input at time step $$t$$, $$h_t$$ is the hidden state, and $$A_t, B_t, C_t$$ are input-dependent dynamic parameters. After Mamba, we apply the inverse GHScan to reconstruct a 2D feature map:
$$F_{\text{mamba}} = \text{GHScan}^{-1}(y) \in \mathbb{R}^{B \times C \times H \times W}.$$
Finally, a Squeeze-and-Excitation (SE) block recalibrates channel-wise responses:
$$\begin{align}
z &= \text{GAP}(F_{\text{mamba}}),\\
s &= \sigma(W_2 \delta(W_1 z)),\\
F_{\text{out}} &= s \odot F_{\text{mamba}},
\end{align}$$
where GAP is global average pooling, $$\delta$$ is ReLU, $$\sigma$$ is sigmoid, and $$W_1, W_2$$ are learnable weights. This space-channel joint modeling (MambaSE) enhances both global context perception and selective channel emphasis, significantly improving small object feature representation for China drone imagery.
2.2 CMambaSE Block
To replace the original C3k2 module in YOLOv11, we design the CMambaSE block, which fuses local convolutional features with global SSM features via a dynamic weighted fusion mechanism. As illustrated in the architecture, the block consists of two parallel branches: the C3k2 branch (local feature extractor) and the MambaSE branch (global context modeler). Let the input be $$X$$. The local features are computed as:
$$F_{\text{local}} = \text{C3k2}(X).$$
The MambaSE branch takes the same input (or optionally the local features) and produces:
$$F_{\text{global}} = \text{MambaSE}(F_{\text{local}}).$$
A lightweight dynamic weight generator, implemented as a two-layer MLP with a global average pooling head, produces a fusion weight $$\alpha$$:
$$\alpha = \sigma(\text{MLP}(\text{GAP}(F_{\text{local}}))).$$
The fused feature is obtained as a weighted sum:
$$F_{\text{fused}} = \gamma_{\text{c3k2}} \cdot \alpha \cdot F_{\text{local}} + \gamma_{\text{mamba}} \cdot (1-\alpha) \cdot F_{\text{global}},$$
where $$\gamma_{\text{c3k2}}$$ and $$\gamma_{\text{mamba}}$$ are learnable branch scaling factors to stabilize training. After a $$1\times1$$ convolution, layer normalization, and SiLU activation, the output is added to the original input via a residual connection with DropPath regularization. The complete forward pass is:
$$Y = X + \text{DropPath}(\text{SiLU}(\text{Conv}_{1\times1}(F_{\text{fused}}))).$$
This dynamic fusion allows the model to adaptively emphasize either local details or global context depending on the input complexity, which is crucial for handling the diverse scales and occlusions found in China drone traffic scenes.
2.3 Optimized FASFFHead Detection Head
Standard detection heads in YOLO series treat different scale features independently, leading to suboptimal performance for small objects. We introduce an optimized FASFFHead that performs spatial-aware multi-scale feature fusion with a small-object bias. The head receives four scale features $$\{X_0, X_1, X_2, X_3\}$$ from the neck. For each output level, the FASFF module fuses three adjacent scales by first resizing them to the same spatial size, then generating spatial attention weights using a lightweight convolutional module:
$$W_i = \text{DWConv}_{3\times3}(\text{Conv}_{1\times1}(X_i)), \quad i \in \{0,1,2\}.$$
The three weight maps are concatenated and passed through a $$1\times1$$ convolution to produce a 3-channel weight tensor, which is then passed through a softmax with a small-object bias. For high-resolution levels (level=2 and level=3), we add a bias vector $$B = [0.6, 0.3, 0.1]$$ before softmax to favor the highest resolution feature. The fused feature is:
$$F_{\text{fused}} = \sum_{i=0}^{2} \hat{W}_i \odot \text{Resized}(X_i),$$
where $$\hat{W}_i$$ is the softmax-normalized weight. The classification branch is enhanced with a deeper structure (six conv layers including depthwise convolutions) for better discriminability. The regression branch uses Distribution Focal Loss (DFL) with increased reg_max from 16 to 20 for finer bounding box localization. The final predictions are:
$$\begin{align}
B_{\text{pred}} &= \text{DFL}(\text{Conv}_{3\times3}(F_{\text{fused}})),\\
C_{\text{pred}} &= \sigma(\text{Classifier}(F_{\text{fused}})).
\end{align}$$
This head significantly reduces missed detections for small targets in dense China drone traffic scenes.
3. Experiments
3.1 Datasets and Settings
We evaluate our model on two public China drone datasets: VisDrone2019 and UAVTrafficControlDataset. VisDrone2019 contains 10,209 high-resolution aerial images with 10 object classes (e.g., pedestrian, car, bus), featuring small objects (over 70% of instances) and diverse urban environments. UAVTrafficControlDataset includes 3,717 images from Lithuanian cities with four classes (Car, Truck, Pedestrian, Bus). All experiments are conducted on an NVIDIA GeForce RTX4060 with PyTorch, using image size 640×640, batch size 2, 150 epochs, AdamW optimizer, and cosine learning rate decay (initial lr 0.001, final lr 0.0001).
3.2 Ablation Studies
We perform ablation experiments on VisDrone2019 by incrementally adding our components to the YOLOv11n baseline. Results are shown in Table 1.
| Config | P | R | mAP50 | mAP50-95 |
|---|---|---|---|---|
| Baseline | 0.492 | 0.189 | 0.343 | 0.206 |
| +CMambaSE | 0.533 | 0.240 | 0.386 | 0.235 |
| +FASFFHead | 0.523 | 0.227 | 0.375 | 0.228 |
| +CMambaSE + FASFFHead | 0.537 | 0.268 | 0.401 | 0.243 |
| +CMambaSE + Optimized FASFFHead | 0.550 | 0.256 | 0.403 | 0.248 |
From Table 1, integrating both CMambaSE and optimized FASFFHead yields the best performance, with mAP50 and mAP50-95 improving by 17.5% and 20.4% over the baseline, respectively. The recall increases from 0.189 to 0.256, indicating better detection of small and occluded objects. Precision also rises to 0.55, confirming reduced false positives.
3.3 Comparison of Fusion Strategies
We compare two fusion mechanisms in CMambaSE: dynamic weighted fusion (our default) and feature adaptive gating. Results are in Table 2.
| Fusion | P | R | mAP50 | mAP50-95 | Params(M) | GFLOPs | Latency(ms) | FPS |
|---|---|---|---|---|---|---|---|---|
| Dynamic Weighted | 0.550 | 0.256 | 0.403 | 0.248 | 7.49 | 26.1 | 33.71 | 29.66 |
| Adaptive Gating | 0.548 | 0.265 | 0.405 | 0.251 | 7.85 | 27.6 | 35.62 | 28.07 |
Dynamic weighted fusion achieves comparable accuracy with lower computational cost (7.49M params vs. 7.85M, 26.1 GFLOPs vs. 27.6 GFLOPs) and faster inference (29.66 FPS vs. 28.07 FPS). Thus, we adopt it as the default in CMambaSE.
3.4 Comparison of Scanning Mechanisms
We compare raster scan and generalized Hilbert scan in MambaSE. Table 3 shows the results.
| Scan | P | R | mAP50 | mAP50-95 | Params(M) | GFLOPs | Latency(ms) | FPS |
|---|---|---|---|---|---|---|---|---|
| Raster | 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 |
Hilbert scan improves precision (0.550 vs. 0.535) and mAP50 (0.403 vs. 0.400), albeit with slightly higher latency due to non-contiguous memory access. The spatial locality preservation helps reduce false positives in China drone scenes.
3.5 Comparison with State-of-the-Art on VisDrone2019
We compare our method with various detectors on VisDrone2019. Results are summarized in Table 4.
| 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 |
| YOLOv6n | 0.517 | 0.168 | 0.343 | 0.209 | 11.9 | 4.2M |
| YOLOv8n | 0.523 | 0.194 | 0.359 | 0.215 | 8.2 | 3.0M |
| YOLOv8-p2 | 0.512 | 0.225 | 0.368 | 0.224 | 12.4 | 2.9M |
| 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 |
| AED-YOLO11 | 0.488 | 0.375 | 0.385 | 0.227 | 12.4 | 2.9M |
| YOLO-S3DT | 0.474 | 0.379 | 0.379 | 0.211 | 11.0 | 2.93M |
| PS-YOLO | 0.454 | 0.339 | 0.323 | 0.185 | 10.0 | 5.0M |
| YOLOv5_mamba | 0.517 | 0.384 | 0.401 | 0.237 | 50.9 | 12.9M |
| YOLOv11n improved | 0.477 | 0.377 | 0.384 | 0.229 | 10.5 | 3.01M |
| Ours | 0.550 | 0.256 | 0.403 | 0.248 | 26.1 | 7.49M |
Our model achieves the highest mAP50 (0.403) and mAP50-95 (0.248) among all methods. While some models like AED-YOLO11 achieve higher recall (0.375 vs. 0.256), our precision is significantly higher (0.550 vs. 0.488), leading to better overall accuracy. The YOLOv5_mamba model, which also uses Mamba, obtains mAP50 of 0.401 but with much higher complexity (12.9M params, 50.9 GFLOPs). Our model strikes an excellent balance between performance and efficiency for China drone applications.
3.6 Comparison on UAVTrafficControlDataset
To further validate generalization, we test on the UAVTrafficControlDataset. Results are in Table 5.
| Model | P | R | mAP50 | mAP50-95 |
|---|---|---|---|---|
| YOLOv5n | 0.894 | 0.794 | 0.821 | 0.614 |
| YOLOv6n | 0.855 | 0.818 | 0.825 | 0.645 |
| YOLOv8n | 0.933 | 0.808 | 0.878 | 0.668 |
| YOLOv8-p2 | 0.883 | 0.822 | 0.859 | 0.644 |
| YOLOv10n | 0.846 | 0.790 | 0.849 | 0.669 |
| YOLOv11n | 0.891 | 0.768 | 0.810 | 0.612 |
| YOLO11s | 0.900 | 0.868 | 0.875 | 0.673 |
| YOLOv5_mamba | 0.870 | 0.818 | 0.833 | 0.624 |
| Ours | 0.886 | 0.892 | 0.909 | 0.685 |
Our model achieves the best mAP50 (0.909) and mAP50-95 (0.685), outperforming the second-best YOLOv8n by 3.1% and 2.5% respectively. The recall of 0.892 is the highest among all methods, confirming our model’s superior ability to detect small and partially occluded targets in China drone traffic scenarios.
3.7 Visualization and Heatmap Analysis
We provide qualitative detection results and heatmaps for typical China drone scenes. Our model (Ours) consistently outperforms YOLOv11n in multi-scale, vertical-view, dense small-object, illumination-change, and nighttime scenarios, with fewer false positives and missed detections. The heatmaps show that our model focuses more accurately on target regions and exhibits stronger background suppression, especially in low-light and complex urban environments.
4. Conclusion
In this work, we presented CMambaSE_YOLOv11, an improved YOLOv11 model specifically designed for China drone traffic object detection. By incorporating the MambaSE module for space-channel joint state space modeling, the CMambaSE block for adaptive fusion of local and global features, and an optimized FASFFHead for multi-scale dynamic fusion with small-object bias, our model effectively addresses the challenges of small objects, scale variations, and background clutter in China drone imagery. Extensive experiments on two benchmark China drone datasets demonstrate significant improvements: mAP50 of 0.403 (17.5% relative improvement) and mAP50-95 of 0.248 (20.4% relative improvement) on VisDrone2019, and mAP50 of 0.909 on UAVTrafficControlDataset. Future work will explore model pruning and knowledge distillation to further reduce computational complexity while maintaining accuracy, enabling real-time deployment on edge devices for China drone-based intelligent transportation systems.