Vehicle pose estimation by parking AGV based on RGBD camera

Parking Automated Guided Vehicles (AGV) continue to play a key role in alleviating urban parking congestion and providing valet parking services. Most of the current parking AGV operate in fixed lanes within stereo garages. The vehicles must be parked in the designated parking area before the parking AGV can complete the vehicle transportation. This places specific demands on the parking driver's skills. When the parking system estimates the vehicle's pose, it typically requires integrating a global camera positioned above the vehicle with sensors mounted on the AGV, including cameras, radars, etc., to achieve the vehicle's pose estimation. This paper proposes a method for the pose estimation of parked vehicles using an RGBD camera installed on a parking AGV. The YOLOv5 network is used to train the model on the RGB images of license plates and wheels, respectively. Subsequently, the generated model file is deployed to the parking AGV. YOLOv5 license plate recognition model detects license plates. Through coordinate conversion, conditional filtering, and point cloud registration, the pose transformation matrix of the vehicle is obtained, and the horizontal rotation angle of the parking AGV is derived. The YOLOv5 wheel recognition model retrieves the bounding box of wheel detection information. By utilizing wheel symmetry, the Euclidean distance between the front and rear wheel coordinate points is computed to ascertain the vehicle's wheelbase and the distance between the parking AGV and the wheels. Subsequently, the distance to park the AGV is calculated. Moving distance. The sensors utilized in this method are solely RGBD cameras, which decrease the number of necessary sensors, improve the maneuverability of the parking AGV, and eliminate track constraints. The proposed method reduces the cost of parking AGV pose estimation and simplifies the detection steps, thereby enhancing detection speed and efficiency.