A nonparametric approach to region-of-interest detection in wide-angle views

• We model motion patterns using distribution of 3D structure tensor-based features.
• We use a nonparametric approach to learn the distribution of features.
• The approach detect both spatial and temporal regions-of-interest in videos.
• Qualitative and quantitative experiments show the effectiveness of the approach.

We propose a tracking-free method to detect the regions of interest (ROI) in a wide-angle video stream. A region is defined as a statistical outlier among occurrences of motion patterns, and is detected in an unsupervised manner. Based on 3D structure tensors, the activity at any site is modeled by the probability distribution of distances between structure tensors. The distribution is estimated using a nonparametric kernel density estimator. The detection of regions is determined by observing a long period of low-probability motion occurrences. Experiments performed with real-world datasets indicate that the proposed algorithm can detect both spatial ROIs and spatio-temporal ROIs, and outperforms other nonparametric methods.