Multiple Vehicle Detection By Improved YOLOv3 In Complex Road Scene

As a crucial part of intelligent transportation system,vehicle target detection technology plays an important role in reducing traffic congestion,reducing traffic accident rate and improving people’s travel efficiency.YOLOv3,one of the most popular algorithms in deep learning target detection,has been widely used in the traffic field.However,due to the complex and variable weather and different degree of vehicle density,its accuracy in vehicle target detection still needs to be improved.To this end,the YOLOv3 algorithm is studied and improved in this dissertation to further enhance its accuracy for the multiple vehicle detection in complex scenes.The main research works are as follows:(1)The YOLOv3 algorithm is improved to address the problems of the low detection accuracy and missed detection.Firstly,to address the problem that the K-means clustering algorithm of YOLOv3 is easy to fall into the local optimal solution in the process of vehicle target detection,this study proposes the method of integrating genetic algorithm and K-means algorithm to improve the way of calculating the distance of K-means by using the feature of genetic algorithm with global advantage,so as to reduce the leakage detection of the model.Then,for the problem that the gradient of the loss function decreases sharply when the prediction frame is close to the real frame,the CIOU loss function is introduced in this thesis to increase the matching degree of the algorithm for the boundary frame,so as to improve the vehicle detection accuracy.Subsequently,to address the problem that the YOLOv3 algorithm is not effective for the small target detection,the MobileNetv3 network is proposed to replace the feature extraction network of the original YOLOv3 algorithm,and the depth-separable convolution is used to replace the convolution module of the original YOLOv3.Finally,it is demonstrated by experimental validation under different weather conditions and dense target complexity: the improved YOLOv3 model promotes the mAP from 76.2% to 82.4% with a little reduction in detection speed compared to the original model.(2)To further promote the attention of the YOLOv3 algorithm to important regions during image generation,an improved model incorporating the attention mechanism is suggested.First,two attention mechanism modules,SENET and CBAM,are analyzed and incorporated into the improved model to suppress irrelevant information in the vehicle target images,thus enhancing the feature extraction capability of the model for vehicle targets and better retaining the effective feature information.Then,the data set of this thesis is processed by data enhancement techniques as a way to strengthen the generalization ability of the model.Afterwards,the results of the ablation experiments show that the CBAM module is more effective in increasing the measurement accuracy of the improved YOLOv3 model compared with the SENET module.At the end,the CBAM-YOLOv3 model is compared with the other vehicle target detection models by setting up comparison experiments.The experimental results show that the introduced attention mechanism module improves the mAP from 82.4% to 85.1% with a small increase in model computation,which verifies the effectiveness of this experiment.