Forward Vehicle Detection Based On Fusion Of Millimeter Wave Radar And Camera Information

Under the condition that automobile market maintains a good prospect and traffic accidents happen frequently in China,ADAS(Advanced Assisted Driving System)has attracted great attention and research from scholars.Target perception,as the lowest and most core technology in ADAS,provides an important basis for the decision-making functions of upper layers.At present,the popular sensors mounted on a vehicle are mainly two types of millimeter-wave radars and optical cameras.Millimeter-wave radars have the characteristics of all-weather,long detection distance,high accuracy but high noise,and lack of target category information.At the same time,cameras have low prices and can rapidly identy the category of target but susceptible to weather conditions.These two sensors have their own benefits and disadvantages.Considring the strengths and weaknesses of the two sensors,this article presents an optimized vehicle detection algorithm based on the cooperation of millimeter wave radar and camera information.The improved algorithm has better accuracy and real-time during vehicle detecting.(1)Effective target determination of millimeter wave radar.After analyzing the basic characteristics and principles of millimeter-wave radar,we choose the appropriate radar for our experiments.Ananlyzing the data measuring by radar and the character of noise to solve the problem of discretization of the original results,a clustering algorithm DBSAN based on density is applied to abtain initial target.Then the Kalman filtering is used to estimate the results of clustering.The thresholds are set to exclude the temporary false targets,and finally the distance,orientation and speed of the effective target are obtained.(2)Implement convolutional neural network to detect vehicles ahead.Convolutional neural network framework is applied to detect the targeting vehicles to solve the problem of low accuracy of traditional visual target detection methods.Based on the analysis of the SSD network structure and training process,the kitti dataset was used to train ssd-tensorflow to generate a model file.Aiming at the problem that the ssd network has a complicated structure and a large amount of calculation,the ssd network is deployed on Movidius,and an embedded deep learning framework is built to realize real-time forward vehicle detection.(3)Data fusion of multi-sensor.The reference frames of millimeter-wave radar,world and the camera are established firstly.Then geometrical relationship of the three reference frame and camera calibration are analyzed to realize the synchronization in space of millimeter-wave radar and visual data.At the same time,the synchronization in time of data is realized by adopting the least common multiple method which solves the problems about the out-sync of frequency collected by sensor.Finally,a decision-level data fusion model is established.The target preliminary selection box is determined based on the radar data.The IOU index is used to verify the preliminary selection box in conjunction with the SSD detection results,and the effective fusion result is output.Through experimental verification,in a road environment with a clear field of view,the method proposed in this paper can effectively detect the position and speed information of vehicles in front of the road,and meets the real-time requirements,and can provide an effective decision basis for the ADAS decision-making layer.