Research On Vehicle Target Detection Algorithm Based On Deep Learning

With the increasingly prominent urban traffic problems,intelligent transportation system has been gradually applied to people’s life.The key part of the intelligent transportation system is the vehicle detection algorithm.The vehicle detection algorithm based on deep learning has gradually become the mainstream detection algorithm for its better detection effect.As one of the commonly used deep learning vehicle detection algorithms,YOLOv3 has a poor detection effect on large and medium-sized vehicle targets,and the detection accuracy of the algorithm on vehicle data sets needs to be further improved.Therefore,this paper proposes an improved YOLOv3 vehicle detection algorithm.The main work of this paper is as follows:Firstly,to solve the problem that the priori box obtained by K-means clustering algorithm does not conform to the size distribution of the real target box in the vehicle data set,this paper uses K-means++clustering algorithm to generate the priori box and improves the distance calculation formula.Then,the feature extraction network based on YOLOv3 algorithm was redesigned to solve the problem of poor feature extraction ability.Finally,aiming at the problem that the detection scale of the feature image of YOLOv3 algorithm is too small,the number of feature images is increased to six,and the size of the six feature images is redesigned.The feature image fusion strategy is used to fuse the feature images of different scales,so as to improve the detection effect of the algorithm for large and medium-sized vehicle targets.To solve the problem that the number of images in KITTI dataset is too small,data enlargement method is needed to expand the dataset.Through comparative experiments,this paper tests the detection accuracy of YOLOv3 algorithm on the Kitti data set processed by several data augmentation methods.The experiment not only verifies that data augmentation can increase the detection accuracy of the model,but also verifies that CutMix data augmentation is the most effective data augmentation method.And the KITTI data set augmented by CutMix data is used as the data set of subsequent experiments.In order to verify the effectiveness of the algorithm improvement strategy in this paper,a comparative experiment was conducted before and after using different clustering algorithms and network structure improvement,proving that the improved strategy in this paper can improve the detection accuracy of the algorithm.In addition,the proposed algorithm is compared with the current mainstream algorithm through experiments,and the detection effect of several algorithms is analyzed.The experiment proves that the proposed algorithm can achieve better detection effect than other algorithms on the basis of guaranteeing real-time detection,and can adapt to the detection task in different scenes.