Human Centered Scene Understanding Based on 3D Long-Term Tracking Data

Scene understanding approaches are mainly based on geometric information, not considering the behavior of humans. The proposed approach introduces a novel human-centric scene understanding approach, based on long-term tracking information. Long-term tracking information is filtered, clustered and areas offering meaningful functionalities for humans are modeled using a kernel density estimation. This approach allows to model walking and sitting areas within an indoor scene without considering any geometric information. Thus, it solely uses continuous and noisy tracking data, acquired from a 3D sensor, monitoring the scene from a bird’s eye view. The proposed approach is evaluated on three different datasets from two application domains (home and office environment), containing more than 180 days of tracking data.