Abstract
In object detection, targets in adverse and complex scenes often have limited information and difficult feature extraction, necessitating deeper feature extraction to adequately capture target features for accurate localization and classification. Addressing the challenge of object detection in complex scenes and low-quality images, this paper designs a lightweight feature extraction network based on CBAM and multi-scale information fusion. Initially, within the Backbone module of YOLOv5s, we construct large-scale, deep feature maps, integrate CBAM, and fuse high-resolution features from shallow networks with deep features. We also add new output heads, employing different feature extraction structures for classification and localization. These improvements significantly enhance detection performance, particularly in challenging scenarios such as strong light, nighttime, and rainy conditions. Experimental results indicate that the improved network structure demonstrates superior detection performance in complex scenes, especially for pedestrian crossing detection under adverse weather and low-light conditions. The study utilizes an open-source pedestrian crossing dataset from Shanghai Jiao Tong University, available on GitHub. Our algorithm improves the pedestrian crossing detection precision (AP0.5:0.95) by 5.9%, reaching 82.3%, while maintaining a detection speed of 44.8 FPS, thereby meeting the stringent requirements of real-time detection. The source code for this program can be found at this address https://github.com/soo-s/yolov5-crosswalk/