Research on YOLO-v3 road target detection based on the combination of K-means++ algorithm and cross-entropy loss function

Abstract

The accurate detection of targets on the road is of great significance (for the big data of road system). However, the traditional detection suffers from missed detection and false detection. In this work, an improved algorithm was proposed based on Yolo-v3 algorithm. The replacement of traditional K-means clustering algorithm with K-means++ algorithm endowed the system a more suitable size of detection frame. Meanwhile, the detection sensitivity of small or incomplete targets was obviously improved with the utilization of cross entropy loss function. The detection experiments on vehicles and pedestrians were performed in two actual situations, normal situation and congested situation. The average detection accuracy was increased by 2.46% with the improvement of clustering algorithm, and the accuracy was increased by 2.08% with the improvement of loss function. The average accuracy was increased by 3.56% with the combination of K-means++ algorithm and cross-entropy loss. The experiments results showed that the combined algorithm possessed a significant improvement in average accuracy, recall and detection speed. This research provides new ideas for the establishment of high-precision algorithms for road dynamic target detection.