Vehicle Detection And Trajectory Tracking Prediction Based On Video Recorder Of Dashcam

The target detection,tracking and trajectory prediction in video is an important research direction in the field of computer vision and vehicle auxiliary driving.However,the current research direction is mainly for video data collected by fixed-position cameras.It is a more meaningful study to accurately detect a specific target under the condition of motion,and track and track it in the corresponding field of view.In this thesis,the video collected by the TripREC on the vehicle is selected as the main research data.Based on this,several key issues of target detection,target tracking and trajectory tracking in dynamic scenes are studied,including: selection of detection models;accurate tracking of moving vehicles;restoration and prediction of running trajectories of running vehicles.In order to solve these problems,this paper proposes corresponding solutions and the specific work is as follows:1.The training model was selected and the standard data set was augmented.The three common target detection networks,Faster-RCNN,SSD,and YOLOv3,was tested,and it was found that YOLOv3 is more suitable for vehicle detection.In order to improve the accuracy of vehicle detection and reduce the rate of missed detection,the VOC2007 + VOC2012 data set was performed for the expansion of vehicle categories only.The new model trained improves the accuracy of vehicle detection.The driving environment such as the weather and background outside would affect the quality of the video captured by the driving recorder.The sunny highway can be used as an ideal driving environment to directly detect the target.After image enhancement on the images of rainy and snowy days,the detection accuracy is significantly improved.2.A target tracking method combining Euclidean distance and Kalman Filter was proposed.According to the similarity measure of Euclidean distance,the change of the detection frame position is measured to determine whether the tracking effect of the same target before and after the change is poor.If the vehicle speed is fast,problems such as target tracking loss are likely to occur.At this time based on the Euclidean distance The Kalman filtering algorithm was added to ensure continuous tracking when the target position changes continuously in vehicle tracking.In addition,extracting lane lines as a filtering signal to screen the target can exclude vehicles that have no impact on the vehicle's driving temporarily,and can reduce the amount of calculation.3.Fit and predict the motion trajectory based on Gaussian fit and vehicle ID.The centroid of the detection frame in the target detection was abstracted as the centroid of the vehicle.By tracking the ID number of the vehicle,the driving trajectory of the corresponding vehicle relative to the driving recorder can be simulated by Gaussian fitting,and the driving in the future period will be based on the previous trajectory Route prediction.At the same time,in accordance with the lane line,corresponding early warning reports are made for irregular driving that is likely to cause safety hazards during driving.The innovations of this paper mainly have the following two points: combining Euclidean distance and Kalman conversion for target tracking,and Gaussian fitting and vehicle ID for fitting and predicting the trajectory.