Songnan Duan

Small-object detection in UAV imagery remains challenging because repeated downsampling weakens fine structural cues, conventional feature pyramids are poorly aligned with small-object-dominated scale distributions, and fixed detection heads provide limited adaptation. To address these issues, a Hierarchical–Bidirectional–Dynamic Network (HBD-Net) is developed as a parameter-compact detector based on YOLOv11s. HEEM reinforces edge, contour, and texture information at multiple backbone stages using orthogonal depthwise strip convolutions, residual detail extraction, and zero-initialized fusion. PLSS replaces the conventional P3–P5 hierarchy with P2–P4 to emphasize high-resolution features, while P2-BiFPN introduces bidirectional propagation and learnable weighted fusion. An integrated Dynamic Head further refines features across scale, spatial, and task dimensions. On VisDrone, HBD-Net achieves 39.0% mAP₅₀ and 22.6% mAP_50–95, improving the YOLOv11s baseline by 6.9 and 4.3 percentage points, respectively, while reducing parameters from 9.4 M to 3.7 M. Ablation studies confirm the complementary effects of the components. On HIT-UAV, HBD-Net reaches 84.1% mAP₅₀ and 57.0% mAP_50–95, indicating effectiveness under infrared aerial imaging. These results demonstrate that coordinated structural enhancement, high-resolution multiscale fusion, and adaptive prediction can improve UAV small-object detection with a compact parameter footprint.

Object detection in unmanned aerial vehicle (UAV) imagery is severely challenged by extremely small object scales, cluttered backgrounds, and pronounced foreground–background imbalance, which jointly degrade the accuracy of general-purpose detectors. This paper presents HSF-Net, a small-object detection network built upon YOLOv11s through three complementary enhancements. First, a Hierarchical Edge Enhancement Module (HEEM) employs orthogonal strip-convolution decomposition with zero-initialized residual fusion to enhance and re-weight the fine-scale edge and texture cues that are progressively attenuated in deep convolutional backbones. Second, a Sparse-Aware Feature Modulation (SAFM) module replaces concatenation-based fusion in the top-down neck pathway, coupling a sparse foreground gate with channel-wise scale modulation to confine cross-scale aggregation to object-bearing regions. Third, the detection head is restructured from {P3, P4, P5} to {P2, P3, P4}, introducing a high-resolution pathway for tiny targets while removing the original low-resolution P5 branch. On VisDrone, HSF-Net attains 46.4% mAP₅₀ and 28.5% mAP_50–95, exceeding the baseline by 7.1 and 4.8 percentage points, respectively, while reducing the parameter count from 9.4 million to 3.6 million. The model achieves an end-to-end throughput of 131.3 FPS on an NVIDIA RTX 3090, although the high-resolution P2 branch increases the computational cost to 39.5 GFLOPs. After dataset-specific training and evaluation on the markedly different TinyPerson benchmark, HSF-Net outperforms YOLOv11s by 5.2 percentage points in mAP₅₀, indicating that its relative performance advantage persists under a substantially different target-scale distribution and maritime background.

Pedestrian trajectory prediction based on vision is a popular task in autopilot system. As pedestrian trajectories always cross each other, pedestrian targets will frequently obscure each other, which makes the collected pedestrian trajectories produce errors. Moreover, the interaction between pedestrians will have an impact on their trajectories in crowded areas, which leads to great challenge in trajectory prediction. In order to solve the above two problems, a pedestrian trajectory prediction method is proposed. Based on the YOLOv7 algorithm, it is connected with the StrongSORT algorithm. Then adding a feature recognition (ReID) module to the model, the problem of target switching and jumping in the process of pedestrian tracking can be solved. The dot product attention mechanism is also integrated into the long short term memory (LSTM) algorithm, the improved LSTM correlate target position and distance information for the pedestrian trajectory prediction. Extensive experiments on three most challenging datasets showed that this method improves the performance and largely reduces the model size.