Visual Odometry Based Vehicle Lane-changing Detection

Lane-changing detection is necessary for accurate positioning, to allow vehicle navigation system to generate more specific path planning. Lane-changing detection method in this paper is more of a deterministic task, proposed based on curve analysis obtained from visual odometry. From the visual odometry trajectory, we have the estimation of vehicle lateral/longitudinal position, yaw, and speed. We also used the road lane information from digital map provided by OpenStreetMap to narrow the lane-changing event possibility. The analysis is conducted on sequences from KITTI dataset that contains lane-changing scenarios to study the potential of lane-changing detection by using visual odometry trajectory curve. Cumulative sum and curve fitting methods were utilized for the lane-changing detection from visual odometry curve. The detection was conducted on several visual odometry approaches for comparison and system feasibility. Our analysis shows that trajectory generated by visual odometry is highly potential for a low-cost and effective lane-changing detection with 90.9% precision and 93.8% recall accuracy to complement more accurate routing service and safety application in Advanced Driver Assistance System.