DetDSHAP: Explainable Object Detection for Uncrewed and Autonomous Drones With Shapley Values

Automatic object detection onboard drones is essential for facilitating autonomous operations. The advent of deep learning techniques has significantly enhanced the efficacy of object detection and recognition systems. However, the implementation of deep networks in real-world operational settings for autonomous decision-making presents several challenges. A primary concern is the lack of transparency in deep learning algorithms, which renders their behaviour unreliable to both practitioners and the general public. Additionally, deep networks often require substantial computational resources, which may not be feasible for many compact portable platforms. This paper aims to address these challenges and promote the integration of deep object detectors in drone applications. We present a novel interpretative framework, DetDSHAP, designed to elucidate the predictions generated by the YOLOv5 detector. Furthermore, we propose utilising the contribution scores derived from our explanatory model as an innovative pruning technique for the YOLOv5 network, thereby achieving enhanced performance while minimising computational demands. Lastly, we provide performance evaluations of our approach demonstrating its efficiency across various datasets, including real data collected from drone-mounted cameras and established public benchmark datasets.