Unmanned aerial vehicles can automate transportation of construction materials and supplies to remote locations. Current operations are performed by manned helicopter crews as the cargo is typically attached with an external sling-load hook and its placement requires precise cooperation between pilot and members on the ground. To ensure the safety of comparable operations with an unmanned rotorcraft, it shall automatically assess the object composition on the ground like a human pilot. Given the limited amount of video data depicting operations in the field from a top-view perspective, a novel dataset of interactions between ground crew, pilot and payload is presented. It includes over 17,000 manual annotations for the aerial object detection of people, ground vehicles, various payloads and cargo hooks. In addition, the crew members are labelled with pose attributes to infer additional state information in the same inference step. Despite the high variation in perspective, a trained detection transformer reached a validation mAP50 of $80.4 \% $, a recall of $76.4\% $ at an inference latency of $50\ ms$ on a NVIDIA Jetson Orin. The dataset is made publicy available under the following link: https://universe.roboflow.com/lft/upsidedownwash.
UDW: A Dataset For Automated Monitoring Of Cargo-Handling During Sling-Load Operations
2023-12-11
1745277 byte
Conference paper
Electronic Resource
English
Automated air-cargo handling systems
British Library Conference Proceedings | 1998
|Airship Sling-Load Operations: A Model Flight-Test Approach
Springer Verlag | 2022
|