A lightweight small object detection algorithm based on improved YOLOv5 for driving scenarios

Small object detection has been a longstanding challenge in the field of object detection, and achieving high detection accuracy is crucial for autonomous driving, especially for small objects. This article focuses on researching small object detection algorithms in driving scenarios. To address the need for higher accuracy and fewer parameters in object detection for autonomous driving, we propose LSD-YOLO, a small object detection algorithm with higher average precision and fewer parameters. Building upon YOLOv5, we fully leverage small-scale feature maps to enhance the network’s detection ability for small objects. Additionally, we introduce a new structure called FasterC3 to reduce the network’s latency and parameter volume. To locate attention regions in complex driving scenarios, we integrate Coordinate Attention and explore multiple solutions to determine the optimal approach. Furthermore, we use a spatial pyramid pooling method called LeakySPPF (Wen and Zhang, in: Jin Z, Jiang Y, Buchmann RA, Bi Y, Ghiran A-M, Ma W (eds.) Knowledge Science, Engineering and Management, pp. 39-46. Springer, Cham, 2023) to further improve network speed, achieving up to 15% faster computation. Finally, to better match driving scenarios, we propose a medium-sized dataset called Cone4k to supplement insufficient categories in the VisDrone dataset. Extensive experiments show that our proposed LSD-YOLO(s) achieves an mAP and F1 score of 24.9 and 48.6, respectively, on the VisDrone2021 dataset, resulting in a 4.6% and 3.6% improvement over YOLOv5(s) while reducing parameter volume by 7.5%.