Vehicle Pose Estimation Based on Computer Vision

Accurate vehicle pose estimation is important for many modern applications, such as autonomous driving, advanced driver assistance systems, and infrastructure based vehicle tracking. In situations where satellite-based systems like GNSS are unreliable, especially in urban environments, roadside sensors can provide a valid alternative. This thesis presents a system that uses two stationary 3D scanners placed at the roadside to estimate the pose of passing vehicles by capturing side views of the wheels. The system is flexible and can be used in various applications, including smart mobility within Internet of Things (IoT) infrastructures, infrastructure-assisted localization, vehicle monitoring, and inspection. One of the main challenges is the placement of the sensors. The scanners must be installed very close to the sidewalk edge to minimize the required space of the system and make it more flexible to different street topologies. This fact reduces the visible portion of the wheel and makes pose estimation difficult. The first part of this work focuses on optimizing the scanner configuration to ensure that a sufficient portion of the wheel is always captured, even at the minimum possible distance from the sidewalk. A simulation tool was developed to test different positions and orientations, and the problem was formulated as a constrained optimization. The second part of the thesis focuses on processing the sensor data. A machine learning model was trained to segment the tire region in the 2D images provided by the scanners. This region is then used to filter the corresponding depth data, isolating the vehicle region of interest from the background. From the filtered point cloud, the system estimates the position and orientation of each wheel. These values are then processed using an estimation algorithm that takes as input the position and orientation measurements. A simulated vehicle model is used to compute the pose, which is combined with the input uncertainty to suppress the non-systematic noise and improve the stability of the final wheel pose estimation. Based on the final wheel positions and orientations, the full vehicle pose and velocity are estimated. The final system is able to estimate vehicle pose accurately and consistently, even with tight installation constraints and a limited field of view. The approach shows strong potential for real-world roadside deployment in a variety of vehicle monitoring and localization tasks.