Research On Vehicle High-precision Collision Warning System Based On Fusion Perception

With the rapid development of autonomous driving technology,increasing road safety and reducing the incidence of traffic accidents has become the focus of the current field of transportation.A high-precision,high accuracy collision warning system is essential for the autonomous driving system.The core of the early warning system is the precise sense of the surrounding environment,and a single sensor has its limitations.The fusion of sensor is the best solution to improve environmental perception ability.The mainstream is to use millimeter wave radar and camera to fusion.In response to the existing fusion,there is no full player of the advantages of the sensor that leads to the discomfort and misunderstanding of the fusion system.In order to perform deeper information complementarity,this article uses millimeter wave radar and camera to detect the location information of the target.The fusion is used as the input of the warning system to complete the warning.In terms of camera detection,the neural network is first based on SSD target detection.The D~2-CITY dataset is used to train and detect the model to deploy the model on the TDA2 S embedded platform.According to the principle of the camera imaging,a single-eye ranging is used to use the point-based reverse vision transformation.The Zhang Zhengyou calibration method is used to calibrate the internal reference of the camera,and optimizes it to the sub-pixel level.For accuracy,for the ranging error that is not allowed to bring in the ranging feature point,it is proposed to optimize the vertical and horizontal positions of the ranging feature point based on gray-based mutations and taillight detection to improve the rangefall accuracy.In response to the impact of changes in the pitch angle on dynamic ranging,the use of the characteristics of the horizontal line is used to compensate for changes in the pitch angle.In order to maintain the proportional information in the resistance of the pitch angle,multi-goal tracking The jumping of the camera rangefinder data is stable with Karman filter.In terms of radar detection,the detection data is obtained by analyzing the message,the stationary target is filtered out according to the relative speed,and the target outside the adjacent lane is filtered out according to the lateral distance,and the radar data is tracked and filtered by using Kalman filter and life cycle algorithm.Preprocessing to remove false targets.In terms of integration,the target detection data obtained by the two sensors,establish space synchronization based on coordinate conversion relationship,realize data synchronization based on the Laglang daily interpolation method,and then combine the targets of the two sensors,and successfully associated the association.Targets are established to reduce the missing and virtual inspection of the sensor.After obtaining the fusion target information,the model of time-based warning is based on the relationship between TTC and the collision time threshold as a judgment condition for warning.Finally,this article will be deployed on the TDA2 S embedded platform.Based on the Vision SDK algorithm framework,the module algorithm LINK is designed.Through testing,the accuracy of the 90.1%was obtained.