Pedestrian detection for underground mine vehicles using thermal imaging

Abstract

Vehicle accidents are one of the major causes of deaths in South African un- derground mines. A computer vision-based pedestrian detection and track- ing system is presented in this research that will assist locomotive drivers in operating their vehicles safer. The detection and tracking system uses a combination of thermal and three-dimensional (3D) imagery for the detec- tion and tracking of people. The developed system uses a segment-classify- track methodology which eliminates computationally expensive multi-scale classi cation. A minimum error thresholding algorithm for segmentation is shown to be e ective in a wide range of environments with temperature up to 26 C and in a 1000 m deep mine. The classi er uses a principle component analysis and support vector classi er to achieve a 95% accuracy and 97% speci city in classifying the segmented images. It is shown that each detec- tion is not independent of the previous but the probability of missing two detections in a row is 0.6%, which is considered acceptably low. The tracker uses the Kinect's structured-light 3D sensor for tracking the identi ed peo- ple. It is shown that the useful range of the Kinect is insu cient to provide timeous warning of a collision. The error in the Kinect depth, measurements increases quadratically with depth resulting in very noisy velocity estimates at longer ranges. The use of the Kinect for the tracker demonstrates the principle of the tracker but due to budgetary constraints the replacement of the Kinect with a long range sensor remains future work.