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

As the future development direction of the automobile industry,intelligent vehicle is an effective solution to problems such as road congestion and traffic accidents.Intelligent vehicle environment perception system,mainly relies on various sensors to accurately obtain the environmental information needed for intelligent vehicle to make decision,it also has become a research hotspot in the automotive field at this stage.The vehicle is one of the most important kinds of targets to be recognized by the environmental information perception system.This paper mainly studies vehicle detection technology based on the information fusion of millimeter wave radar and machine vision.Firstly,this article analyzes the current research status of vehicle detection technology at home and abroad,and compares the advantages and disadvantages of various sensors,the main research content of this article is clarified by summarizing the deficiencies of existing vehicle detection technology.Secondly,based on the analysis of causes of invalid signals in raw radar data,the radar data preprocessing method for setting effective area and hierarchical clustering is identified.for special situations such as radar works unstably,the consistency check strategy is used for analyzing the existence situation of the target,and the life cycle method is used for studying the entire existence process of the target.Based on predefined rules,the selection of vehicle targets is completed.Thirdly,the data set suitable for domestic vehicle detection is made by combining the KITTI data set and the self-made sample set.The sample set is used for training and analyzing the vehicle detection performance of YOLO v2 network.The real-time performance and detection ability for small targets of network is improved by tailoring network structure properly and adding multi-scale detection,therefore,the improved YOLO v2-mini network model is got.Fourthly,a spatial fusion model is established based on coordinate transformation and sensor calibration,and the thread synchronization method is used for achieving sensor time fusion based on the principle of downward compatibility.The region of interest of visual detection is confirmed according to the effective targets of radar,and YOLO v2-mini is used for processing region of interest to get visual detection results.The sensor fusion strategy is established based on the intersection of union of detection frames and Global Nearest Neighbor data association method,the sensor information fusion can be achieved by processing the detection results of radar and visual with the sensor fusion strategy,and the Extended Kalman Filter is used for tracking fusion targets to complete the decision of the fusion target.Finally,the experimental platform is built relied on the intelligent vehicle of the research group,each sensor is calibrated.The real vehicle experimental is conducted under various road conditions,the experimental results verify the functionality and effectiveness of the sensor information fusion strategy proposed in this paper.