Pedestrian Localization and Trajectory Reconstruction in a Surveillance Camera Network
Hai An Vu, Van Giap Nguyen, Anh-Tuan Pham, Thanh-Hai Tran · 2017
In this paper, we propose a high accuracy solution for locating pedestrians from video streams in a surveillance camera network. For each camera, we formulate the vision-based localization service as detecting foot-points of pedestrians in the ground plane. We address two critical issues that strongly affect the foot-point's detection results: casting shadows and pruning detection results due to occlusion. For the first issue, we adopt a removing shadow technique based on a learning-based approach. For the second issue, a regression model is proposed to prune the wrong foot-point detection results. The regression model plays a role in estimating the position by using the human factors such as height, width and its ratio. A correlation of the detected foot-points and the results estimated from the regression model is examined. Once a foot-point is missed due to uncorrelated problem, a Kalman filter is deployed to predict the current location. To link the trajectory of the human in the camera network, we base on an observation about the same ground-plane/floor in view of cameras then the transformation between a pair of cameras could be computed offline. In the experiments, a high accuracy performance for locating the pedestrians and a real-time computation are achieved.