Research on real-time detection of road vehicle targets based on YOLOV4 improved algorithm

Abstract

Aiming at the problem of low accuracy of single-step detection algorithm and long time-consuming detection of two-step detection algorithm, this paper proposes a convolutional network model that has the characteristics of fast single-step detection algorithm and high accuracy of two-step detection algorithm. It can meet the requirements of real-time detection of road vehicle targets. The algorithm proposed in this paper is based on the end-to-end network yolov4, by deepening the number of layers of the convolutional network to obtain more semantic features. At the same time, in combination with the anchor mechanism, the K-means++ clustering algorithm is used to cluster out suitable vehicle widths and heights. The benchmark frame uses four feature acquisition frames, and then uses FPN+PAN to perform feature fusion, and finally uses CIOU_Loss as the loss function of coordinate prediction to obtain a more excellent algorithm. Through comparison and analysis of the results, the algorithm in this paper is significantly improved compared to other similar algorithms.