An Application of Object Detection Based on YOLOv3 in Traffic
Sujin Luo, Chenyu Xu, Hongxin Li · 2019
With the progress of scientific-technology, automatic equipment has become a prevailing trend in our society. Especially in transportation, unmanned vehicle system is a significant research project that can greatly benefit us. In traffic scene, to achieve high detection accuracy in real-time, a robust algorithm is needed. In this paper, we present a detection method based on YOLOv3 which preprocesses the data set before training. We divide the original images into equal parts with k-fold cross validation. Each time when we train the model, we choose one part as testing set and remaining parts as training set to make full use of our data. The experiment results show that our method performs better than YOLOv3 in the case of class-imbalance. We also propose a new network structure, Dense-YOLOv3 which replaces some residual modules with dense modules. Through this method, overfitting is effectively alleviated.