ABSTRACT
As the world's population has grown and mobility has become more accessible, human activities have expanded into remote areas and deserts. These activities (typically desert safaris) generate much left-behind waste. Untreated left-out wastes contribute directly to gas emissions that exacerbate the global warming, especially litter such as plastics, cans, or any other resistant material. We present a drone-based solution to find waste in extra-urban environments (mainly deserts). In this paper, we use an autonomous unmanned aerial vehicle (drone) that flies over a desert area to capture a video stream to locate and identify waste objects such as plastic bottles. Our solution is still a work in progress; hence, in this paper, we have explored an object detection algorithm using transfer learning (few-shot-learning), where a pre-trained classifier (YOLO) is retrained with our relatively small dataset DroneTrashNet. Our goal is to benchmark existing models and approaches for object detection and classification in order to provide a baseline for selecting the most appropriate models for detection and classification tasks in extra-urban contexts. Our DroneTrashNet dataset corresponds to three minutes of annotated video captured by a drone at a low altitude in a controlled experiment in a desert environment, where waste objects were manually placed, with some objects half-buried in the sand. Preliminary results show an achieved mAP of 42% in specific conditions that could be subject to a number of improvements. The first results are promising compared to the results obtained with the initial classifier. While the resulting model has been validated in a controlled desert environment, new challenges and opportunities remain to be addressed, such as full-scale validation and generalization to extra-urban environments.
