Research On Vehicle Detection And Tracking Technology Based On Radar And Camera Fusion

The vehicle-road cooperative system has a great potential to improve the efficiency of road passage and reduce the accident rate,and therefore became a critical part of the national transportation industry construction.With the great potential to improve the efficiency of road passage and reduce the accident rate,the vehicle-road cooperative system has become an important part of the national transportation industry construction.As the core of the vehicle-road cooperative system,the road perception undertakes the important task of using multi-sensor fusion to detect road information and realize the fast understanding of dynamic traffic environment.As an important part of road sensing technology,vehicle detection and tracking technology aims to use intelligent devices to collect road information for intelligent analysis in order to provide a decision basis for traffic management.The main research content of this thesis is how to design intelligent algorithms to accurately sense and track vehicles on real roads by using multi-sensor information.Some researchers have already conducted research on vehicle detection and tracking technology under multi-sensor.However,the existing research work has the following shortcomings: 1)the existing traditional multi-sensor fusion detection method has a high leakage detection phenomenon,which cannot meet the demand of high leakage detection requirements for scenarios such as vehicle-road cooperation;2)the existing traditional multi-sensor fusion tracking method cannot solve the problem of the uncertain number of vehicle targets and the sudden appearance or disappearance of vehicle targets.(1)Radar and camera spatio-temporal calibration methodRadar and camera alignment principles in time and space are studied in this thesis.In the temporal dimension,the proposed alignment method,which takes the millimeter-wave radar sampling period as the fusion moment and combines multi-threading technology,completes the temporal alignment with the collected camera data;in the spatial dimension,the detailed derivation of the conversion formula for mapping the radar data under the roadside to the pixel coordinate system is completed,and the spatial conversion between the radar and the camera data will be completed;the real camera calibration parameters are obtained by Zhang Zhengyou’s method,which lays the foundation for studying the vehicle detection and tracking method of radar and camera fusion.(2)Method of vehicle detection based on fusion of coarse and fine level of resolution This thesis investigates the fusion method of vision based on background modeling with radar.A fusion detection method with a coarse and fine level of granularity is proposed to correct the fused vehicle detection frames by combining the gradient histogram and support vector machine techniques for the vehicle detection phenomenon.Compared with the background modeling-based method and the radar fusion method,the proposed method can improve and reduce the correct detection rate and the missed detection rate by 2.48%,respectively;compared with the background modeling-based vision method,the proposed method can improve and reduce the correct detection rate and the missed detection rate by 12.35%.(3)Method of vehicle tracking based on PMBM filterThis thesis investigates the tracking algorithm based on the PMBM filter.In the multi-sensor framework,regarding the miscorrelation problem between radar and camera data in the traditional data correlation method,the fusion correlation strategy of radar and camera is proposed in combination with the global nearest neighbor method,which reduces the miscorrelation problem when correlating radar and camera data;regarding the traditional tracking methods which are difficult to deal with the unknown number of targets and the sudden appearance or disappearance of targets,the fusion association strategy is combined with the PMBM filter algorithm to propose a PMBM vehicle tracking method based on the fusion association strategy of radar and camera,to support the solution of the target tracking problem under the unknown number of vehicles and the sudden appearance or disappearance of vehicle targets effectively.According to the experimental validation,the proposed method reduces the GOSPA value by 34.96 on average compared with the GM-PHD tracking algorithm under the GOSPA metric,effectively improving the robustness of vehicle tracking.In summary,this thesis combines real cameras as well as radar hardware devices to design and implement a multi-sensor fusion-based road monitoring system.Results of the experiment showed that the proposed method in this thesis can effectively detect and track vehicles using radar and camera fusion information,validating the viability of the method proposed in this thesis.