A monovision-based 3D pose estimation system for vehicle behavior prediction

Collision accidents have become a crucial and urgent problem. Advanced Driver Assistance Systems (ADAS) can enhance traffic safety and improve efficiency. Toward this direction, we propose a robust low-cost system based on monovision to estimate the accurate 3D pose information of the target vehicles for further behavior prediction. Firstly, we combine the Haar features with AdaBoost classification for vehicle detection, and then use the Hypothesis Verification to remove the false detection; Then, the system keeps tracking and locates the license plate(LP); Finally, the typical PnP problem is applied to estimate the 3D pose information including the rotation matrix and translation matrix relative to the target vehicles. Compared to stereovision systems, the monovision-based system significantly reduces the computational cost and has a lower requirement for equipment. Moreover, it can work on slope road and is easily applied to different countries due to the standardized LP in each country. The experiments in the real-world have been tested, showing excellent performance in vehicle detection, LP location and 3D pose estimation.