Automated Height Adjustment for Truck-Trailer Coupling

Abstract

Coupling a truck to a trailer is a common but non-trivial task in logistics, typically requiring manual height adjustment to align the truck’s fifth wheel with the trailer. This thesis explores an automated approach using computer vision to estimate the relative height between the truck’s fifth wheel and the trailer, and automatically adjust the height accordingly. The inputs available for estimating this height is the camera feed from a singular backup camera mounted at the rear of the truck, and sensor data from the trucks air suspension.

The proposed system performs keypoint detection, locating corners, utilizing the YOLOv11 pose estimation model. We explore two different methods for calculating distance and then utilize this information to calculate the height difference between the trailer and the fifth wheel of the truck. The first method is based on assuming the width of the trailer and counts the number of pixels between the two bottom corners. The second method assumes both the width and the height of the trailer and solves the perspective-n-point problem to obtain the distance.

The model was tested in real-world scenarios and achieved high keypoint precision and recall. The distance calculation remained within 10% of the ground truth most of the time, which is adequate for this application. Furthermore, the error in height estimation was within 7-15 cm of margin of error which was sufficient to enable automatic height adjustment for the truck-trailer coupling.

The results indicate that the full system is successful on most of the trailers tested. This confirms the practical feasibility of using keypoint detection to understand the geometrical relationships between the camera and the trailer it is observing, which could also be useful for other assisted driving applications. Our results indicate that more training data are needed to obtain robust height and distance estimates when approaching a trailer from the side. The constraints our model must place on the shape of the trailer it is detecting also highlight the need for research into other methods with weaker assumptions about the shape of the trailer.