Edge AI-Driven Video Analytics: A Mobile Deep Learning Framework for Victim Detection in SAR Robotics

摘要

In post-disaster search and rescue (SAR) missions, it is crucial for robots to distinguish between actual victims and dummy objects, despite their similar characteristics. Edge video analytics have demonstrated exceptional performance in such missions, however, the trade-off between high-performing deep learning models and computational complexity often results in inadequate computational efficiency for SAR robotics. This study tackles this challenge by deploying a deep learning (DL) algorithm, i.e., MobileNet Single Shot Detector (SSD), specifically tailored for embedded systems, to perform object detection and classification tasks on lab-scale SAR robots. The proposed algorithm is developed in Python and implemented on a low-powered edge computing device. The effectiveness of the proposed algorithm in object localization is highlighted by its mAP score of 0.6247 and an average IoU of 0.9036. Experimental tests using a confusion matrix showed that the model accurately predicted 42 dummy objects and 43 victim dolls. The model achieved an overall accuracy of 89.47%, with a precision of 1.0, a recall of 0.8947, and an F1 Score of 0.9444. Real-time testing confirmed the model's robust performance, maintaining high accuracy and a stable frame rate of 47 FPS. These findings corroborate the efficacy of the MobileNet SSD algorithm in reliably detecting and differentiating objects within embedded SAR robotic operations.