Application Research Of Vehicle Detection Technology Based On Fusion Of Millimeter Wave Radar And Vision

With the rapid increase of vehicle ownership year by year due to the development of China’s national economy,traffic congestion and accidents are inevitably caused.The government is actively constructing a smart transportation system to alleviate the traffic pressure.Real-time monitoring of road and vehicle conditions is the key to building a smart transportation system.Currently,the vehicle detection methods used in traffic are generally divided into four types: induction coil,radar,geomagnetic,and vision.Induction coil detection and geomagnetic detection are generally used for parking lots or road checkpoints due to their detection principle limitations.Radar detection can accurately detect the position and speed of the target and is less affected by external environmental changes,but its recognition rate decreases when the vehicle is occluded.Vision detection can effectively obtain target categories on the road,but it is susceptible to environmental changes.Therefore,combining the advantages of radar detection and vision detection and using the fusion method of millimeter-wave radar and vision system for vehicle detection is an effective method for real-time vehicle detection.Firstly,YOLOv5 s algorithm is chosen as the target detection algorithm used by the vision system for vehicle detection.To address the problems of large parameters and slow inference speed of YOLOv5 s algorithm on embedded platforms,Shuffle Netv2 and SENet attention mechanism,both lightweight neural networks,are introduced to improve the YOLOv5 s algorithm.The improved algorithm is named Shuffle-YOLOv5s-SE.Images of vehicles are collected at multiple intersections and labeled as datasets for algorithm training.YOLOv5 s and Shuffle-YOLOv5s-SE are trained and verified under the same experimental environment.The results show that Shuffle-YOLOv5s-SE algorithm reduces the parameter size by 45% and FLOPs by 47.9% compared to YOLOv5 s algorithm on computer platforms.The inference speed on both CPU and GPU is improved.Both algorithms are deployed on the RV1126 embedded platform.The results show that the memory usage of Shuffle-YOLOv5s-SE algorithm is reduced by 31.4% compared to YOLOv5 s algorithm when inferring images on the embedded platform.The inference speed of each frame is increased by 5ms.Secondly,the principle of measuring static and dynamic targets by millimeter-wave radar is analyzed,and the type of millimeter-wave radar is selected.The target information detected by the millimeter-wave radar is parsed based on the CAN communication protocol.Then,the coordinate system transformation is applied to the output data of the millimeter-wave radar and the vision system to achieve spatial fusion of the two sensors.Suitable sampling frequency is selected to achieve time fusion of the two sensors with synchronous sampling.Considering the actual traffic conditions,the Kalman filtering algorithm is used to track the targets detected by the vision system continuously,in order to reduce the occurrence of special cases of misfusion by introducing motion feature matching.Finally,a millimeter-wave radar and vision fusion experimental platform is built.The proposed vehicle detection method based on millimeter-wave radar and vision fusion is verified to have a 2.7% improvement in recognition rate and a 1.7% reduction in false detection rate compared to the single vision detection method in well-lit traffic scenes.In poorly lit traffic scenes,the proposed method has a 13.2% improvement in recognition rate and a 2.5% reduction in false detection rate compared to the single vision detection method.