Lane-changing Detection And Tracking For Preceding Vehicle With On-board Driving Video

Drivers usually change lanes for the purpose of overtaking and avoiding obstacles on raod.While improper operations such as illegal line pressing behavior could interfere with the operation of rear traffic,which is potentially dangerous in traffic-intensive or fast-speed sections and may even lead to collisions with vehicles in adjacent lanes.It is necessary to supervise risky lane changing behaviors accurately and efficiently.As a key component of big data technology,machine vision could use unstructured image and video data to perceive various traffic phenomena,which provides effective solutions for the above task.Since most traffic monitoring devices have limited monitoring vision,a large number of vehicle monitoring devices on road could be used as data sources to capture and track lane-changing behavior of the preceding vehicle from a mobile perspective,providing rich information for risk assessment.This can not only be applied to the field of traffic law enforcement,but also form crowdsourcing mechanism to help build a safe and efficient road traffic system.In this paper,the vehicle on-board driving video become one of data sources for realizing the perception of traffic operation information ahead,through which a detection and tracking mechanism is built for lane-changing and line-pressing behavior of preceding vehicle.The risk assessment is performed based on the acquired multi-dimensional data and unsupervised learning algorithm.First,for the detection and tracking module in information perception,this paper improves feature map extraction structure in Mask R-CNN model and applies it to the preceding vehicle detection according to research needs.Based on the detection results,continuous tracking of preceding vehicle is achieved combined with the vehicle motion and appearance characteristics between video frames.This paper mainly applies the Unscented Kalman Filter(UKF)prediction to correlate vehicle motion features in different frames and proposes a sub-domain detection method based on re-identification network to achieve appearance feature correlation.In addition,this paper realizes lane line detection and classification respectively through cascaded convolutional neural networks.Redundant pixel removal module is added after the detection network to improve semantic segmentation performance.Then,the perceived pixel coordinates of vehicles are converted into real world coordinates through the on-board camera calibration,this helps achieve preliminary information extraction through integration.This paper designs rules based on the perception results,which directly determines the line pressing state of preceding vehicle in the original picture.It records and binds information of preceding vehicles in each historical frame.Regarding the ID number generated by vehicle tracking as an index,this research captures the start and end moments of the line-pressing behavior with 5 frames as a sliding time window,through proposed information backtracking method.The research takes these two moments as the boundary and further determines the start and end moments of lane-changing process in the preceding and following frames,through vehicle running informationThe research uses the information backtracking method to obtain coordinates of lane changing process.Multidimensional metrics are further extended and refined by computation after preprocessing.The data of each frame contains six dimensional features after integration,including minimum vehicle distance,lateral speed,lateral acceleration,type of pressure line,weather conditions,and light intensity.All frame information is pooled together to form a sample dataset which will be clustered according to the set number of clusters respectively through the Gaussian Mixture Model GMM.The Calinski-Harabaz(CH)index is used as the clustering performance evaluation standard.The study selects the cluster center corresponding to the maximum CH index to classify the risk levels of various lane-changing behaviors to realize risk assessment.After the case video test,the accuracy of the lane change detection and tracking mechanism reaches 91.2%.In the clustering process,the CH index reaches the maximum when the number of clusters is five.Finally,the research classifies the risk level of lane changing behaviors in the video to realize risk assessment.