Pedestrian detection and trajectory estimation in the compressed domain using thermal images

Since a few decades, the Unmanned Aerial Vehicles (UAVs) are considered precious tools for different military applications such as the automatic surveillance in outdoor environments. Nevertheless, the onboard implementation of image and video processing techniques poses many challenges like the high computational cost and the high bandwidth requirements, especially on low-performance processing platforms like small or medium UAVs. A fast and efficient framework for pedestrian detection and trajectory estimation for outdoor surveillance using thermal images is presented in this paper. First, the detection process is based on a conjunction between contrast enhancement techniques and saliency maps as a hotspot detector, on Discrete Chebychev Moments (DCM) as a global image content descriptor and on a linear Support Vector Machine (SVM) as a classifier. Second, raw H.264/AVC compressed video streams with limited computational overhead are exploited to estimate the trajectories of the detected pedestrians. In order to simulate suspicious events, six different scenarios were carried out and filmed using a thermal camera. The obtained results show the effectiveness and the low computational requirements of the proposed framework which make it suitable for real-time applications and onboard implementation.