Optimized Yolov5s Algorithm for Enhanced Autonomous Driving Obstacle Detection

This paper presents a video object detection method based on the Yolov5s model, primarily designed for obstacle detection in autonomous driving scenarios. The Yolov5s model was optimized by introducing the Ghost Net network, ECA attention module, Mixup data augmentation method, Bi-LSTM network, KF filter, and ConvLSTM visual memory module. Additionally, a balance between traditional feature extractors and extremely lightweight feature extractors was implemented. The combination of these improvements not only leverages temporal information in the video stream but also addresses spatiotemporal correlation and performs feature fusion through the ConvLSTM visual memory module. This enables the Yolov5s model to handle obstacle detection tasks in autonomous driving scenarios with higher accuracy and efficiency. Experimental results show that the improved Yolov5s network model has a weight file size of 11.4MB on the ImageNet VID2017 dataset, an FPS of 72.6, a Multi-Object Tracking Precision (MOTP) of 0.81, and a Multi-Object Tracking Accuracy (MOTA) of 0.74. Compared to other video object detection methods, the Yolov5s-based approach demonstrates higher accuracy, continuity, and stability in video object detection and tracking, and can effectively handle obstacle detection challenges in complex environments