ABSTRACT To address the challenges of false negatives and false positives of small objects and the difficulty of fine‐grained behavior recognition in complex traffic scenarios, this paper constructs a hybrid deep learning framework based on image detection to synergistically improve multi‐object localization accuracy and semantic understanding capabilities. The framework first uses a combination of Gaussian and bilateral filtering for denoising, enhancing input quality and improving detection sensitivity for small objects. In the detection phase, the YOLOv5s (You Only Look Once 5s) model is used as the baseline. The Convolutional Block Attention Module (CBAM) attention mechanism is applied to enhance the representation of key features. K‐means clustering is used to adaptively generate prior anchor boxes that match the scale distribution of objects in traffic scenarios. The CIoU (Complete Intersection over Union) loss function is also used to optimize bounding box regression accuracy, improving small object detection performance while maintaining model lightweight. To achieve fine‐grained semantic understanding, a two‐branch classification network is designed. The attribute branch uses the ConvNeXt‐Tiny (Convolutional Next‐Generation Tiny) structure to extract static appearance features, while the event branch utilizes the nonlocal operations module to capture dynamic contextual dependencies. Weighted fusion of these two features enables joint recognition of attributes and behaviors. A GNN‐CNN (Graph Neural Network‐Convolutional Neural Network) hybrid classification module is also constructed. The GNN models the spatiotemporal interactions between vehicles, while a lightweight CNN extracts local texture features. These features are adaptively fused using the Squeeze‐and‐Excitation (SE) attention mechanism, and a softmax classifier performs traffic behavior judgment. Experiments show that the YOLOv5s‐CBAM model achieves a mean average precision (mAP) of 0.55 for detecting extremely small objects (< 16 × 16). In the overloaded vehicle detection task, the GNN‐CNN module achieves accuracy and recall of 0.92 and 0.90, respectively. This hybrid deep learning framework provides reliable technical support for automated traffic inspections. It improves the accuracy and stability of small object detection and fine‐grained event recognition in complex traffic scenarios. Its modular design and strong scalability make it widely applicable and conducive to promoting intelligent transportation towards higher levels of automation.
Chen et al. (Wed,) studied this question.