Automatically detecting and tracking people walking through a transparent door with vision

There has been a growth in demand for surveillance equipment to monitor people in indoor as well as outdoor environments. Furthermore, using guards to watch surveillance screens all the time is highly inefficient and thus automation of human monitoring can be more accurate and produce cost savings. The problem is challenging if we choose to use a passive non-invasive sensor such as vision. The specific problem we investigate is tracking people through a sliding glass door. This is challenging because of the transparent door and both the door and person are moving. The method we have chosen consists of tracking coherent motion field clusters. The video frames are preprocessed, corner features are extracted and matched over frames, and the background trajectories are learnt. Finally, the test sequences are processed to obtain the trajectories of the various image features and those are classified based on the background model into foreground and background trajectories. The proposed method was tested on a set of real data with varying scenarios, and illumination as well as noise changes with a success rate approaching 95% correct classification into either background or foreground even if the tracker lost track of the entering person.