Improved YOLO Algorithm for Object Detection in Traffic Video

Abstract

The detection of moving targets in complex traffic scenes is the most basic and important technical means in video surveillance. In order to balance the speed and accuracy of object detection, this paper chooses You Only Look Once(YOLO) algorithm to extract foreground targets in video frames. Meanwhile, some steps are used to improve this algorithm. First, data is augmented during the pre-processing phase to ameliorate the imbalance of sample category. Then, re-clustering our own data set before training to get the corresponding anchor box size to enhance the accuracy of the final training model. In the training process, the focus loss is used instead of the binary cross entropy loss to further solve the problem of slow convergence rate and poor training effect caused by the imbalance of sample category. The improved YOLO algorithm is used to compare the training results with the original YOLO algorithm, and they are comprehensively analyzed by the model evaluation index. It can be verified that the improved YOLO algorithm maintains a faster training speed while it also improves the accuracy of training.