SNCE-YOLO: An Improved Target Detection Algorithm in Complex Road Scenes

Abstract

Autonomous driving is a prominent research area, with the primary challenge being the precise perception and interpretation of the surrounding environment. To address issues such as low resolution, dense occlusion, and small targets in complex road scenes, we propose an improved road target detection algorithm, SNCE-YOLO, based on YOLOv5. Firstly, to avoid the loss of feature information during training for low-resolution and small targets, the Space-to-Depth Convolution(SPD-Conv) is introduced to replace the original convolutional layer for retaining global feature information. Secondly, to address the issue of confusing complex backgrounds with detection targets when features are extracted too early and deeply, the Normalization-based Attention Module (NAM) is incorporated into the backbone network without increasing the computational effort, focusing on the channel and spatial information of interest. Thirdly, to enhance the capture and fusion of multi-level feature information, we optimized the neck network's feature fusion module by implementing the C2f structure. Finally, to reduce the bias in the regression of the occluded target positions, we introduced the EIOU loss function to better define the relationship between the predicted and real frames. The SNCE-YOLO algorithm achieves a mean Average Precision (mAP@0.5) of 91.9% on the KITTI dataset, which represents a 2.6% improvement over the baseline model. Additionally, with a detection speed of 95.24 frames per second (FPS), it meets the requirements for real-time target detection in complex road scenes. Comparisons are also made with other mainstream target detection algorithms to verify the effectiveness of SNCE-YOLO.