Research On Preceding Vehicle Detection Method Based On Millimeter Wave Radar And Deep Learning Visual Information Fusion

The multi-source information detection technology of the preceding vehicle is one of the important research directions in the field of Advanced Driving Assistant System(ADAS)and driverless systems.In the currently used millimeter-wave radar and visual information fusion method,the static millimeter wave radar has a narrow beam angle and the vehicle detection of traditional machine vision has long time processing and low precision of detection,which makes it difficult to accurately monitor and track the curved road in real time.The preceding vehicle target(hereinafter referred to as the tracking target)causes the loss of the tracking target and even causes the vehicle rear-end collision.Therefore,aiming at the problem of millimeterwave radar tracking target loss and traditional machine vision vehicle detection with high timeconsuming and low precision during cornering,a preceding vehicle detection method based on millimeter-wave radar and deep learning visual information fusion is proposed.Firstly,the second chapter deals with the loss of the tracking target of millimeter wave radar during cornering.According to the millimeter wave radar ranging speed measurement principle,the millimeter wave radar detection model of tracking targets on different driving lanes(straight and curved)is established.This paper analyzes the relationship between the radius of the curve and the state of self-driving when the tracking target is lost,and proposes a millimeter-wave radar and visual camera information fusion method for tracking target loss.The method is to track the target millimeter wave radar historical data.The information is state information,and the visual camera image information of the tracking target is used as measurement information,and the driving state parameters such as the distance and speed of the tracking target are calculated by using Unscented Kalman Filter(UKF),and used as the next time tracking.The state information of the target is continuously tracked to solve the problem that the tracking target of the millimeter wave radar is lost when the curve is traveling.Secondly,the third chapter is based on the high time-consuming and low-precision problem brought by the traditional machine vision vehicle detection method for visual cameras.The YOLOV3 deep learning algorithm based on convolutional neural network is used to detect the tracking target.In order to speed up the detection speed of the algorithm,the YOLOv3 algorithm model is simplified by using the pruning quantization method,and the training of the YOLOv3-tiny algorithm model is completed,so that the real-time and accurate tracking target detection requirements are achieved in the detection time and detection precision.Then,the fourth chapter combines the above two aspects,establishes the data fusion model of millimeter wave radar and visual camera,and completes the spatial coordinate uniformity of millimeter wave radar and visual camera according to the coordinate conversion relationship,and completes the millimeter according to the minimum common sampling period.The time coordinates of the wave radar and the vision camera are unified,thus establishing a front vehicle detection system based on millimeter wave radar and deep learning visual information fusion(hereinafter referred to as fusion system);in this fusion system,using millimeter wave radar to generate a sense of tracking target The interest area is verified by the YOLOv3-tiny deep learning algorithm,and the merged tracking target is calculated according to the established fusion rules.Finally,the UKF algorithm is used to realize the real-time tracking of the tracking target.Finally,Chapter 5 verifies the effectiveness of the fusion system tracking target detection method.The selection of each component of the fusion system was carried out,and the fusion system consisting of millimeter wave radar and visual camera was installed and debugged.The tracking target status information was obtained by using the system and the VBOX detection device based on high-precision GPS mutual positioning.The tracking target detection experiment under different curvature radius of curvature is completed.By comparing and analyzing the VBOX data and the tracking target data detected by the fusion system,it is verified that the curve tracking target loss algorithm effectively solves the problem of corner target loss.