Pedestrian detection model in complex environment based on improved YOLOv5s

For solving the problems of high false detection rate and missed detection rate of pedestrian detection in a complex environment, an improved model YOLOv5s-RFDH based on Yolov5s is proposed.The model retains the YOLOv5s baseline network while optimizing and improving the feature extraction and detection parts to enhance the accuracy and robustness of pedestrian detection in complex scenes.This paper focuses on pedestrian detection in the CrowdHuman and WiderPerson datasets, which contain dense crowds and substantial occlusions. To address this challenge, the K-Means++ clustering algorithm is used to re-cluster the datasets and obtain anchor boxes suitable for the data. The Receptive Field Block (RFB) module is introduced for feature extraction, using dilated convolutions in different branches to increase the receptive field and extract deeper feature information. These features are then fused to improve the detection accuracy of small pedestrian targets. Decoupled heads can solve the scale invariance problem in object detection. By introducing a decoupled detection head that separates the classification and regression tasks, targets of different scales and sizes can be accurately detected.Experimental results on the test sets of CrowdHuman and WiderPerson datasets show that the improved model achieves an increase in detection accuracy and a decrease in missing rate. Specifically, the detection accuracy is improved by 1.4% and 1.2% on the respective datasets, while the missing rate is reduced by 2.0% and 1.7%. These results demonstrate the effectiveness of the proposed algorithm.