Adaptive Thresholding for Small Drone Obstacle Avoidance Using Monocular Depth Estimation

Engineering notes

Key topics: autonomous driving, depth estimation, monocular depth, monocular camera. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Recently, the demand of drones has increased dramatically due to low costs and high convenience. Drones are employed in variety fields, e.g., border security, search and rescue, surveying and recreational activities. Drones have become indispensable tools in each of these fields. In particular, small drones (weighing less than 100 grams) are suitable for operation in tight indoor spaces owing to their compact form factor and flexibility in movement. Furthermore, in Japan, the law requires registration or permits for flying drones weighing 100 grams or more, leading to increased expectations for the utilization of small drones in the industry. However, small drones face the challenge of being unable to carry high-performance sensors due to strict weight limitations. Specifically, realizing autonomous flight requires a depth sensor to measure distances to obstacles, but many small drones cannot be equipped with one. To address this issue, this study proposes an obstacle avoidance method using a monocular camera. In this method, depth images are obtained from monocular camera images using a monocular depth estimation model, and then binarized into regions that either contain nearby obstacles or do not. Next, the largest inscribed rectangle in the region where no obstacles are present is determined, and the drone is directed to fly toward the center of this rectangle to avoid obstacles. Evaluation in a simulation environment confirmed that the proposed method achieved up to 39% higher obstacle avoidance success rate compared to existing methods.

Links and sources

• Official / arXiv page

Comments