Real-time human detection and tracking in infrared video feed

Abstract

Human detection is an attractive field in computer vision. It remains a challenging task in constrained light conditions due to low contrast against the background. Most of the research using visible light cameras have mainly focused on human detection when sufficient light is available. They have used RGB images, which were highly sensitive to changes in visible light conditions. This paper is focused on the infrared video feed, which is less sensitive against visible light illumination changes compared to RGB images. We propose a novel approach to human detection using Deep Convolutional Neural Networks (DCNN) and Kernelized Correlation Filters (KCF). The proposed methodology utilizes the MobileNet, pre-trained on ImageNet for frame-wise human detection and KCF to improve the tracking accuracy with motion dynamics model. In addition to that, preprocessing on raw images has been applied to reduce the noise effects. The proposed methodology performed well in infrared image sequence with 89.71% of human detection accuracy, which is higher than the state of the art methods and real-time applicable in currently available surveillance systems.