Research On Vehicle Detection Method In Complex Road Scene

Vehicle detection,as a research branch of object detection,has been widely used in intelligent transportation systems,automatic driving,intelligent parking,and traffic accident handling.In recent years,as domestic and foreign scholars have applied deep learning to object detection,many object detection algorithms based on deep learning have appeared and have shown good detection performance.For vehicle detection,in complex road scenarios,due to the factors such as large changes of the vehicle scales,partial occlusion and light changes,reducing the missing detection rate and false detection rate of vehicle detection still faces severe challenges.There is still a lot of research space in designing reasonable network structure in order to effectively extracting object feature and constructing appropriate loss functions to improve model performance,which is of great benefit for improving vehicle detection performance.Therefore,this thesis takes the general object detection model as the starting point,based on a detailed analysis of the mainstream object detection algorithms,and researches on large-scale changes of the vehicles and partial occlusion in vehicle detection.The main research contents are as follows:1.Improved the SSD algorithm.According to the vehicle detection task in complex road scenarios,aiming at the defects in SSD algorithm: the low detection accuracy of multi-scale vehicles and the missing detection under partial occlusion,we retain the multi-scale detection idea and modify the network structure to improve the feature expression ability of the network.The specific modifications are as follows: using cascaded dilated convolution on the output of specific layers of the backbone network to resolve multi-scale changes in vehicles and adopting dense connections between dilated convolutions to obtain object context information to solve the problem of object partial occlusion.Experimental results show that the improved algorithm's m AP reaches 89% in the UA-DETRAC dataset and it is robust to vehicle scale changes,partial occlusion and light changes.2.Engineering vehicle detection.Aiming at the problem of imbalanced data types in engineering vehicle detection applications,focal loss function was introduced.By adjusting the loss weights of the positive and negative samples in the training samples,the positive and negative samples in the data tend to balance,thereby improving the detection accuracy of the model.The experimental results on the special vehicle dataset show that this method not only effectively solves the problem of imbalanced data categories,but also improves the detection accuracy of the algorithm,and it's m AP reaches 90.6%.