Research On Indoor People Counting Technique Based On Image And Depth Information

Nowadays,with the development of science and technology,the world population has been increasing rapidly,which is a great challenge to social management.For better resources allocation,people counting based on surveillance video can be utilized for transportation planning,goods stockpiling and people mobility in certain events.For indoor scenarios,there exists many challenges including the change of luminance,people gesture,size and occlusion problem,which require for high robustness and real-time algorithm.However,traditional people counting technique is based on a single camera,which leads to inevitable spatial information loss in transformation from real 3D world to 2D images.As a result,this paper invests significant effort into taking maximum advantage of the available depth information and focuses on how to use all available scene information to realize people counting technique.In this paper we use binocular camera for indoor scenario recording,estimate initial depth information by belief propagation algorithm and optimize the result by combining with plane-fitting algorithm.Then,we get the point cloud through depth information and extract region-of-interest using scenario labelling algorithm to improve adaptability for various scenarios.Furthermore,two-level cascade method is used for region-of-interest to reduce false positive and improve the robustness of the algorithm.For the better result in object detection,we propose a local maximum suppression searching strategy to reduce the number of sliding detection window and improve the accuracy of algorithm.Finally,we employ a simplified version of the robust multi-hypothesis tracking based on Extended Kalman Filter and extend it by using the geometric scene context as spacetime constraint.