Research And Application Of YOLOv3 Algorithm In The Field Of Intelligent Transportation

Object detection is one of the important research topics in the field of computer vision,and its core task is to detect the objects of interest in static images or dynamic videos.Driven by technologies such as big data,smart chips,and the Internet of Things,deep learning-based object detection models have gradually replaced traditional detection models,and have been widely used in self-driving,video surveillance and other intelligent transportation fields.In order to perceive the road environment in real time,feedback the traffic target information in time,and provide a reference for the next decision,the system needs a fast and highly accurate detection model to complete the classification and positioning tasks.In this paper,after the video stream collected from the road is processed into an image,it is input into two classical models,Faster R-CNN and YOLOv3 for detection.Then the detection accuracy and speed are compared and analyzed.Finally,YOLOv3 with better performance is selected as intelligent transportation Detection algorithm in the field,and carry out related research work on the basis of it.The main achievements of this paper are as follows:(1)According to the characteristics of the traffic object data set constructed in this paper,The K-means clustering algorithm with improved distance measurement is used to regenerate the Anchors(priori box)that are more in line with the sample size of this paper,and assign them to the prediction layer of the corresponding scale.It provides more accurate position information for the border regression of the model.(2)Combining the ideas of YOLOv3 and DenseNet model,a detection model based on deep feature fusion(YOLOv3-Dense)is proposed.By changing part of the residual structure in the pre-basic network Darknet-53 of YOLOv3,the model fuses coarse and fine granularity features with dense connections instead of skip connections,so as to realize the interactive reuse of deep and shallow feature information,effectively improve the traffic object detection performance of the model.The experimental analysis shows that the mAP(mean average precision)of YOLOv3-Dense reaches 78.47%,which is 4.1%higher than that of YOLOv3,and the detection rate of single-class object can be increased by 14.5%.(3)Focal Loss is used to improve the classification cross-entropy loss of the model.By adaptively allocating weights to positive and negative samples,difficult and easy samples in the training process,the model pays more attention to the study of a small number of difficult samples,effectively reduces the impact of unbalanced class distribution,optimizes the training process,and further improves the overall detection performance of the model.Experiment shows that the mAP value of this method reaches 77.86%,which is 3.48%higher than that of the original model,in which the AP(average precision)value of bus is as high as 97.34%,and the improvement rate is more than 20%.