Object depth measurement based on monocular vision and point transformation for unmanned aerial vehicles

Accurate depth measurement is critical for obstacle avoidance in autonomous unmanned aerial vehicles (UAVs). Vision sensors are attractive for UAVs due to their low cost, compact size, and low weight. This paper presents a monocular vision–based depth measurement method that uses point-based transformation to estimate obstacle depth. An ideal depth measurement model is first established, where depth is inferred from changes in the lengths of corresponding feature line-segments between adjacent frames. To address non-ideal conditions with UAV attitude angles, a point-based transformation is introduced to correct the image line-segment lengths. In the proposed framework, YOLOv5 is used for object detection and SIFT descriptors are employed to extract feature points, which are then used to form feature line-segments. Compared with line-based correction methods, the proposed method eliminates the theoretical error introduced by attitude angles and improves measurement accuracy. Unlike deep-learning-based depth estimation, it does not require pixel-level depth-annotated datasets and only relies on sufficient texture. Simulations and laboratory experiments show that the proposed method achieves a depth measurement error of approximately 8.9% for objects, although it has not yet been validated in real flight scenarios.