Multi-person pose tracking with occlusion solving using motion models

Abstract

We present a method for the multi-person human tracking problem including occlusion solving. To track and associate frame-by-frame human detections obtained using a deep learning approach, we propose to combine motion features extracted by optical flow and Kalman filtering, which allow us to predict the future poses of targets. By taking advantage of the characteristics of both motions features, we are able to handle sharp motions of the target and occlusions. With our simple occlusion handling mechanism, we achieve comparable results with state of the art and are able to keep track of a target identity even when occlusions occur.