Scene understanding is a challenging task and and mainly based on geometric or object centered approaches. Hence, the aim of this paper to introduce a novel human centered approach for scene analysis and tackle challenges of noisy long-term tracking data obtained by a depth sensor. Hence, fast filtering mechanisms are proposed to filter noisy tracking data, reducing the number of outliers and thus significantly improving the accuracy of the detection of walking and sitting areas within indoor environments. Evaluation is performed on two different scenes containing 18 and 34 days of tracking data and shows that detecting and filtering invalid tracking information dramatically increases the accuracy.
展开▼
