Research And Application Of Vehicle Re-identification Algorithm For Parking Lot

With the the improvement of people’s living standards,urban roads and transportation become more and more prosperous at once,it’s difficult to satisfy the actual parking demand just rely on the special parking lot.As a result,many cities meet the parking needs of citizens by dividing vehicle parking area on both sides of some by-pass.However,the management of such parking lots requires a lot of human and material resources,which brings huge challenges to urban management.The vehicle reidentification algorithm studied in this thesis is applied to this scene to realize the intelligent management of parking lots on both sides of the road.However,in the research of vehicle re-identification algorithm based on road scene,in addition to being limited by the limited resolution of high-level cameras,complex traffic conditions and changing weather conditions,etc.,in the public vehicle reidentification data set and actual road scenes,the change of vehicle pose,the interference of vehicle taillights on the appearance of the vehicle and the possible partial occlusion will bring great difficulties to solving the problem of vehicle re-identification.In order to capture the local features of vehicle,I redefine the vehicle key point detection problem and designs a vehicle key point detection algorithm VKPAODN to locate the local feature area of the vehicle in this thesis,which can complete the three tasks of vehicle key point coordinate detection,vehicle direction prediction and vehicle key point visibility judge.Then in order to allow the model to capture more robust vehicle features under limited conditions,I proposed a network with three different branches called MGNVKP in this thesis.MGNVKP focus on the overall features of the vehicle appearance,the vehicle local features based on vehicle key points,and the the vehicle structure features obtained from different vehicle orientations.The overall features and structural features share the Res Net50 feature extractor for feature extraction,and the local features are extracted from the vehicle key points located by the vehicle key point detection algorithm proposed in this thesis.Corresponding attention mechanisms are also designed for the structural feature branch and the local feature branch to capture more discriminative features in MGNVKP.In the end,in this thesis,the model is training and testing on the Ve Ri776 vehicle re-identification data set,the accuracy is up to 79.5%,reach the leading level.Besides,the inference speed of a single picture on GTX 2080 ti is 26 ms,and the inference speed of a single picture on Jeston TX2 embedded platform is 150 ms.In order to strengthen the validity and practicability of the model,I optimizes the problems of taillight light interference in this thesis,spatiotemporal information utilization,and occlusion processing in actual scenes,and verifies the effectiveness of the optimization measures.Finally,MGNVKP is deployed on the embedded platform Jeston TX2 based on tensorrt reasoning engine.Under the model accuracy of FP32,the reasoning speed on GTX2080 ti is about 7ms and that on Jeston TX2 is about 35.6ms,which is enough to meet the real-time use requirements of the system.optimizes and improves the original intelligent parking system through the vehicle re-identification algorithm MGNVKP,which improves the overall accuracy of the intelligent parking system from 65.31% to 79.38% in average in this thesis.