SPR-YOLO: A Traffic Flow Detection Algorithm for Fuzzy Scenarios

Efficient and highly accurate traffic vehicle detection plays a crucial role in intelligent transport. However, in ambiguous scenes such as night and rainy days, factors such as noise interference and low resolution often limit the detection effect. Therefore, this paper proposes a lightweight network architecture for fuzzy scenarios, SPR-YOLO. The model is based on YOLOv8, the backbone and neck modules of the lightweight network are redesigned, and SPD_Conv is adopted to mine deeper semantic information to face the feature extraction in fuzzy scenarios. Task. In order to further enhance the feature aggregation ability of the model, we propose the SECA attention module, which improves the model’s ability to focus on the information in both channel and spatial dimensions for better extraction of semantic features. In addition, in order to achieve high fine-grained fusion effects even in low-resolution and blurred scenes, we propose the DY_GELAN aggregation network to achieve high-fidelity fusion and low-parameter balancing, which further enhances the network’s ability to express deep information. Finally, we use ByteTracker for vehicle tracking and a target statistics method with customized regions to achieve traffic flow detection in fuzzy scenarios. The network is trained and evaluated on the UA-DETRAC dataset. The results show that the parameters of the proposed network architecture are basically at the same level as YOLOv8, but the mAP50 and FPS are improved by 6.4% and 7.68%, respectively. Compared with other mainstream models, the proposed model effectively balances the advantages of lightweight, efficiency and high accuracy.