Research On Object Detection Based On Millimeter-wave Radar And Vision Fusion For Autonimous

With the booming development of traditional manufacturing industry and new era information technology,autonomous vehicles have gradually entered people’s life.In order to ensure that autonomous vehicles can make timely and safe driving decisions,accurate obstacle detection research in complex traffic scenes has attracted wide attention,which brings new challenges to object detection technology in the field of computer vision.Traditional two-dimensional object detection relies on deep learning technology to achieve good results in monocular images.However,the traffic scene also needs to consider environmental factors such as weather,so the perception task of complex traffic environment cannot be completed only by relying on the camera.The effect of 3D object detection on monocular images is poor,mainly because the camera has no ranging function and monocular images lack depth information.Therefore,the current 3D detection researches are mostly based on Lidar.However,Lidar is also limited by weather factors,and has a small sensing range and high cost.Millimeter wave radar through electromagnetic wave to achieve ranging and speed measurement function,and has strong penetration,strong anti-interference ability,all-weather work,and the camera has complementary advantages.Therefore,the sensor fusion strategy based on millimeter wave radar and vision can achieve performance improvement in both two-dimensional and three-dimensional object detection tasks.To sum up,the main research content of this paper is the object detection technology based on millimeter wave radar and vision integration.The main innovations and contributions involved are as follows:(1)A two-dimensional object detection method based on adaptive multi-strategy information fusion is proposed.Firstly,a radar data enhancement method is designed.Aiming at the problem of sparse features of current millimeter-wave radar data,the influence range of each radar point is extended by bilateral filtering,and the feature richness of radar image is enhanced.Then,an adaptive fusion network suitable for radar features is designed.The model can fuse radar and image features at multiple levels and learn the fusion weight independently.Finally,on the basis of feature-level fusion,decision-level fusion is added to realize multi-strategy information fusion.After the open source data set test,the proposed model has obvious performance improvement in the two-dimensional detection task.(2)A 3D object detection method based on radar area suggestion network and depth enhancement is proposed.In this work,an area suggestion network based on millimeter-wave radar points is first designed to obtain better initial anchor frame.Then,a depth feature enhancement module based on bilateral filtering is designed to enhance image features with millimeter wave radar ranging information to make up for the lack of depth information in monocular images.After testing,the proposed model has better performance than before in 3D object detection.