A Method for Estimating Vehicle Heading Deviation and Lateral Position Deviation by Combining Deep Learning and Kalman Filtering

The path tracking control technology for unmanned vehicles based on visual perception has become increasingly mature in structured road scenarios. However, in unstructured roads, due to the significant increase in uncertainties in external environments and vehicle motion states, relevant research is still far from mature. The main issues faced by pure image perception algorithms in unstructured road scenarios include: the variability of external environments, significant impact of roads on vehicle motion states, which greatly affects image quality, leading to inaccurate vehicle state estimation and even abnormal jumps in the estimated values, severely affecting path tracking accuracy and driving safety. To address these issues, this paper designs a Kalman filter post-processing algorithm based on pure image perception algorithms using deep learning. This algorithm integrates the heading angle deviation and lateral position deviation estimated from images with the current actual vehicle speed. Simulation results indicate that the proposed method can effectively reduce estimation errors and suppress abnormal jumps in the estimated values. The algorithm was applied to real vehicle control on a grassland dirt road, and within a speed range of 0 to 20 km/h, the system operated stably, with normal road tracking, maintaining lateral position deviation within 0.6 meters and heading angle deviation within 0.05 radians.