Real-time Safety Monitoring Vision System for Linemen in Buckets Using Spatio-temporal Inference

초록

Linemen risk falls, electric shocks, burns, and other injuries during the daily job and these incidents can often be fatal. In this paper, we present a novel vision-based real-time system for detection and tracking of various non-rigid safety wearables worn by linemen, in a highly cluttered environment. We set up four imaging sensors on the repair truck's bucket to robustly monitor the linemen from four different viewpoints. In the monitoring system, we firstly apply a novel fast background segmentation method to suppress false positives and reduce search space. Next, we represent each safety wearable with a Gaussian mixture model and track them with an LK-tracker. In order to track occluded or out-of-camera-view safety wearables, we propose a novel human pose inference method. The proposed method is an extension from the existing CNN-based human pose inference by utilizing light-weight color, shape, and space-based human pose inference mechanism. The proposed human pose inference method shows improved performance in terms of precision, recall, and speed. Experimental results on a number of challenging sequences demonstrate the effectiveness of the proposed scheme, under complex background, prolonged occlusions, and varying color, shape, and lighting.

키워드