RS-YOLO: An efficient object detection algorithm for road scenes

In response to the demands for high detection accuracy and real-time performance in road scene object detection algorithms, as well as the challenges posed by complex background interference, diverse object categories, varying scales, and occlusion, this paper proposes an effective road scene object detection algorithm RS-YOLO based on YOLOv8s. Firstly, to improve RS-YOLO's ability to represent and perceive semantic features of different levels, an MS-PAFPN (Multi-Scale Path Aggregation Feature Pyramid Network) is proposed. Secondly, to enhance the algorithm's ability of inter-layer feature interaction and intra-layer feature regulation, a WS-Fusion (Weighted Shuffling Fusion) is proposed. Furthermore, a SPPFormer (Spatial Pyramid Pooling Former) is proposed, as a Transformer-like spatial pyramid pooling module, leveraging MaxPool and MLPBlock to improve the algorithm's ability to re-extract semantic features from local details and global context information. The experimental results demonstrate that RS-YOLO achieves highly competitive performances in the SODA10M and KITTI benchmarks, with mAP50 reaching 67.79 % and 95.61 %, respectively, and excellent performance of 48.19 % and 78.51 % on mAP50:95. Compared with YOLOv8s, RS-YOLO's mAP50 improves by 3.7 % and 0.58 %, and its mAP50:95 increases by 4.83 % and 2.23 %, respectively. This achievement provides valuable technical support and solutions for the development and application of autonomous driving technology, as well as new ideas for improving the performance of object detection algorithms in other scenes.