LIGHTWEIGHT TARGET DETECTION ALGORITHM IN COMPLEX TRAFFIC SCENES

Aimed at the problems of large model size and imbalance between detection speed and detection accuracy of traditional target detection algorithm in complex traffic scenes, a lightweight YOLO algorithm based on YOLOv4 is proposed. The lightweight network Mobilenetv1 was used to replace the feature extraction network of YOLOv4, and a cross stage partial module DW-CSP was proposed to reduce the learning of redundant information. At the same time, a new Lish activation function was designed, K-means + + clustering algorithm was used to generate new priori boxes, and the FocalLoss function was introduced to alleviate the imbalance between positive and negative samples. Experiments on specific dataset show that the improved YOLO algorithm has significantly improved the detection speed, detection accuracy and model size compared with the YOLOv4 algorithm, and the effectiveness of the improved algorithm is proved.