| As the number of urban vehicles increases,the capacity of closed parking lots cannot meet the actual demand.Building an on-street parking system through computer vision technology has become an effective solution.However,the urban road traffic situation is complicated and the resolution of the monitoring camera is not high.This leads to problems such as vehicle occlusion and motion blur in the screen,which affects the recognition of vehicle ID,color,and type attributes.Therefore,thesis designs and implements a vehicle attribute recognition algorithm that integrates local features and applies it in actual scenarios.Firstly,in response to the problem of difficulty in extracting vehicle appearance features due to vehicle occlusion,thesis designs and implements a vehicle key point detection algorithm.The algorithm detects unobstructed vehicle key point areas,completes vehicle key point localization,predicts vehicle key point visibility,and predicts vehicle direction.The backbone of the network uses HRnet V1 to extract vehicle key point features and adds an IBN structure to improve it.Considering the interference of background information and occlusion on vehicle key point localization in actual scenarios,thesis introduces CBAM and RGA attention modules in the network to focus on obtaining detailed information and global structural information,and suppressing background information.Simultaneously using unbiased coordinate system transformation and DARK decoding methods,ensuring the accuracy of key point coordinates during image transformation operations and key point decoding.This article evaluates the algorithm in the vehicle key point detection dataset,with a key point accuracy of 96.81%,key point visibility accuracy of 97.01%,and direction prediction accuracy of 86.59%,achieving a relatively leading level.Then,based on the vehicle key point detection algorithm,Thesis designs and implements a vehicle attribute recognition network that fuses local information in a multitask manner.The network adopts a two-branch structure to extract the local features of the key points of the vehicle and the overall features of the vehicle.In order to obtain more representative fusion features,thesis designs and implements a multi-scale attention feature fusion module,and uses the attention mechanism to adaptively calculate fusion weights on different scales to fuse local vehicle features and overall vehicle features.Finally,through the attribute recognition network,the vehicle ID,color,and type attributes are identified.In thesis,a field dataset is constructed under the on-street parking scene,and the attribute recognition algorithm is evaluated on the public dataset Veri776 and the field dataset.On the public data set Veri776,the ID,color,and type recognition accuracy of the model are 80.7%,96.79%,and 95.83%,respectively,reaching a relatively leading level.On the field data set,the model also showed good results.Finally,thesis deploys the designed and implemented vehicle attribute recognition algorithm to the Jetson TX2 platform.A new vehicle matching algorithm is designed and realized by using the vehicle attribute recognition algorithm,improve the vehicle matching module in the roadside parking system,enhance system performance,and design a vehicle matching result visualization software for management personnel to analyze. |
