Research On Vehicle Detection Method Based On Deep Learning In Traffic Scenarios

For many years,in the research process of computer vision,the process of discussing and exploring the target detection algorithm has been going on continuously,and object detection has gradually become one of the most important research topics in the deep learning field.It can be seen from the research status at home and abroad that the application field of target detection is very wide,such as the detection of pedestrians,vehicles and road signs;the detection of vegetable varieties and the quality of the fruit;for the defect detection of equipment parts in industrial machinery.Target detection has been slowly applied in all aspects of life,however,there is more or less development space in the various target detection algorithms that have been researched.First,the accuracy and speed of detection are the two most important evaluation indicators in the target detection algorithm,but they are hard to achieve the optimal balance in the actual design process.Based on the in-depth analysis of the current mainstream target detection algorithms,we research the application of the yolov3 and yolov4 in pedestrian and vehicle detection.The main research work is as follows:(1)Aiming at the problem of yolov3’s insufficient ability to detect small target vehicles in traffic scenarios,we have improved the structure of the yolov3’s network model,and replace the five DBL components in the model with two residual structures and one DBL component to enhance the model’s feature detection ability for small target objects;we optimized the loss function of yolov3,improved width and height loss function and target center coordinate loss function,and use the normalization method to increase the model’s attention to the center coordinates of small-scale targets;at last,we use the bicubic interpolation method to unify the scale of the input image,enhanced image data and maintain the higher resolution of the image.We verified the effect of the new model on the KITTI and Pascal VOC 2007 data sets,the m AP of the new model reached 90.85%,and the new model’s ability to detect small targets has been greatly improved.(2)In addition,we also researched and improved the yolov4’s model.First,we add the deep separable convolution structure to the backbone network of yolov4,greatly reduces the network parameters of the model,reduces the cost of training,and improves the training speed;then,we use the swish activation function to replace the mish activation(3)function in the yolov4’s backbone network and add nonlinear factors to the model;we retainthe SPP structure and PANet structure of yolov4,the SPP structure can more effectively increase the receiving range of the main features,and the PANet structure can more closely integrate the feature information.The new model effectively solves the problem of yolov4’s missed detection in a complex environment,and the detection accuracy and detection speed of the model have also been improved.We have made a series of improvement measures on the yolo’s series of algorithms and verified them in large data sets and actual application scenarios.The improved algorithm model has a good detection effect.