Research On Forward Vehicle Detection Algorithm Based On Fusion Of Vision And Millimeter Wave Radar

Perception is an extremely important part whether in an unmanned system.Accurate and real-time detection of the vehicle can provide sufficient prior conditions for subsequent decision-making and planning,which the system can avoid traffic accidents.Millimeter wave radar and camera have their own advantages and disadvantages in vehicle detection.Millimeter wave radar can effectively obtain the relative distance,relative speed and relative angle of the vehicle,but can’t have accurate information in height of the target.The camera can obtain the target’s width,category,but it can’t accurately measure the distance and speed of the target.There may exist wrong objects when using a single radar or a single camera.Based on the above analysis,this paper proposes a vehicle detection algorithm based on millimeter wave radar and visual fusion.The target point of millimeter wave radar is projected onto the image through time-space transformation,and the association between the radar target sequence and the camera target sequence is realized by the target correlation algorithm.More complete in formation is output for the successful target,and the unrelated target also exit in system.Compared with the single radar or single vision algorithm,more complete road target information can be obtained.The specific research contents are as follows:1.Acquiring the effective vehicle target sequence of radar.Firstly,the Kvaser CAN is used to receive the radar target signal.By analyzing the characteristics of the hollow target,stationary target and invalid target signal of the echo signal,the radar effective vehicle target acquisition algorithm is proposed.The target signal is filtered out by analyzing the features of the radar,the invalid target signal is filtered out by the threshold between the two frames,the refined target signal is filtered b y the relative speed relationship between the vehicle and the target,and finally the effective vehicle target information is realized.2.Acquiring the effective vehicle target sequence of camera.Firstly,introduce the principle of convolutional neural network and traditional target detection algorithm.Then propose the vehicle detection algorithm based on SSD(Single Shot multibox Detector).Secondly,the KITTI data set is preprocessing as trained sample data,and the network model is built and trained through CAFFE.Finally,the robustness of the model is verified by the KITTI test set and the real road sample test set.3.Measuring the vehicle distance based on monocular vision.Firstly,the principle and deficiency of the traditional monocular vision ranging algorithm are analyzed,and proposing the distance measurement method which based on pixel fitting.The relationship between the pixel value and the real distance of the multiple targets in the image is obtained through camera calibration.Finally,using the lower boundary of bounding box of to calculate the actual distance.4.Time-space synchronization of millimeter wave radar and camera.Firstly,the relationship between the pixel coordinate system and the world coordinate system is established by calibrating the camera.Secondly,establish the relationship between the radar coordinate system and the world coordinate system by radar calibration.The radar point is projected onto the pixel coordinate system to complete the spatial synchronization between millimeter wave radar and camera.Finally,using multi-threading to achieve time synchronization between millimeter wave radar and camera.5.Associate millimeter wave radar target and camera target.Firstly,analyzing the target information of millimeter-wave radar and camera,then design cost function by using the multi-feature between the radar target and the camera target to correlation6.Design conduct platform and vehicle detection system.According to the algorithm scheme of this paper,the hardware architecture and software architecture design of the experimental platform are carried out,and the radar algorithm,visual algorithm and multi-sensor fusion algorithm are transplanted into the system.Experiments show that the proposed algorithm performs well.