Research On Vehicle Detection Based On Deep Learning

Target detection is a basic problem in computer vision,while deep learning has obtained widespread application in the testing scene.However,due to various interference factors,such as detecting target angle changes,blocking each other,limitation of carrying equipment and computing power,resulting in the low detection accuracy in the actual scene.With the application of target detection used more widely,the demand for target detection speed and precision are increased.This thesis mainly researches vehicle detection based on deep learning,which focus on fast vehicle detection problem with lightweight equipment and multi-view vehicle detection problem in practical scenes respectively.Corresponding algorithms are developed and systems are deployed,and the research work in vehicle detection is extended to practical scenes for verification.The main contributions include the following three points:1.The thesis proposes a lightweight vehicle detection method based on central point regression,which studies fast vehicle detection on lightweight equipment platform.Specifically,a feature extraction detection mechanism combining residual network and Center Net to simplify the network structure and improve the detection speed.The depth separable convolution is used to enlarge the receptive field without reducing the resolution,and reduce parameters and computation.The feature fusion module is introduced in the up-sampling operation,and the effect of feature extraction is improved by fully considering low-level spatial information and high-level semantic information.Meanwhile,a new matching mechanism is designed and the loss function is improved to further improve the detection accuracy.Simulation results show that the detection speed of the proposed method can reach 191 fps,recall rate can reach 99.0%,and accuracy can reach 99.7% on lightweight equipment.2.The thesis proposes a multi-view vehicle detection method based on weighted Hough voting,researching high-precision detection in actual scenes.Specifically,the thesis improves the process of generation,matching and voting detection calculation of visual words in Hough voting.Meanwhile,the thesis has integrated enhanced random forest into image block clustering,redefined visual words with stronger expression ability and more compact,and improved the accuracy of matching.Furthermore,a weighted detection mechanism is proposed to share information among the sub-classes of each perspective,and the weights of different perspectives are allocated by voting combination weights to improve the accuracy of detection.Moreover,a subclass classification mechanism is proposed to realize the weight of the voting combination and improve the recognition effect of the network as a whole.Simulation results show that the proposed method can effectively improve the detection accuracy of vehicles from multiple perspectives,with recall rate reaching 81.5% and accuracy rate reaching 79.0%.3.The thesis develops a vehicle detection and recognition system based on deep learning,which applies the improved algorithm to specific scenarios.The system takes personal computer as the client and cloud computing host as the server,which effectively supports vehicle detection and recognition requirements in different scenarios.The thesis shows the detailed design,research and development process and test results of the system,which has proved the value of the research and practical application.