A Method for Abnormal Trajectory Detection and Pre-Correction Based on Vehicle-Borne Images

Objectives: Trajectory data obtained from low-cost crowdsourcing devices often suffer from overall positional offsets and abnormal attribute recordings, failing to meet the accuracy requirements for high-definition map updates. Given that onboard images in crowdsourcing data can capture changes in the driving environment and vehicle orientation, this paper aims to explore correction methods for overall trajectory position offsets and abnormal driving angles based on continuous onboard images. Our goal is to provide high-quality trajectory data that enhances the accuracy of subsequent position correction methods. Methods: This study first employs structurefrom-motion technology to estimate the pose of the onboard camera, followed by the detection and correction of abnormal driving angles based on the estimated pose information. Next, the study identifies key trajectory points located at intersections and other critical locations using semantic information extracted from onboard images. The corrected angle values and hidden Markov model are then utilized to match these key trajectory points with nodes, enabling the detection and correction of overall trajectory position offsets. Finally, Wuhan city is selected as the research area, and 30 trajectory data accompanied by onboard images are used to validate the effectiveness of the proposed method. Results: Experimental results show that: (1) The difference between the estimated driving angle and the true driving angle of the vehicle is within 30°, meeting the accuracy requirements of map-matching algorithms. (2) The overall position correction accuracy of the trajectory is within 30 meters, with potential sub-meter accuracy at best. This performance is primarily influenced by the quality of the road network data, highlighting that the method relies on high-quality road network data to effectively address abnormal trajectory data with significant positional offsets. (3) High-quality trajectory data obtained after pre-correction can significantly enhance the accuracy of map-matching algorithms, resolving issues related to incorrect matches and relative positional deviations. Conclusions: The proposed innovative method effectively detects and corrects abnormal driving angles and overall trajectory offsets, significantly enhancing the quality of crowdsourced trajectory data. This improvement further aids in increasing the accuracy of trajectory position correction in map-matching algorithms. Additionally, it offers valuable technical support and methodological guidance for advancing the application of high-definition map crowdsourcing updates.