A YOLOv5s-based Method for Intelligent Pedestrian Detection in Vehicles

This research aims to enhance the efficacy of object detection methodologies in complex traffic scenarios. The focus is on introducing an intelligent approach for identifying vehicles and pedestrians, leveraging the YOLOv5s framework. The study delves into the YOLOv5s architecture, implementing improvements by substituting the BottleNeck CSP module with the Ghost-BottleNeck module to decrease the computational burden. Additionally, the loss function is refined by incorporating Alpha-IoU as the localization loss function, effectively minimizing the discrepancy between predicted and actual bounding boxes. Furthermore, from an inference perspective, the SAHI strategy is introduced to enhance the inference process, significantly improving the detection performance of pedestrians. Finally, ablation experiments and real traffic road scene trials confirm that the SDYOLOv5 pedestrian detection method not only sustains commendable real-time performance but also elevates accuracy in pedestrian detection. This research holds significant implications for advancing intelligent vehicle testing and detection.