Spatio-Temporal Consistency to Detect and Segment Carried Objects

We present a new method to detect carried objects and to segment them accurately after detection. The proposed method includes several contributions: first, a new superpixelbased descriptor is proposed to identify carried object-like candidate regions using human shape modelling. Second, integrating spatio-temporal information of candidate regions to detect carried objects. We exploit the consistency of recurring carried object candidates viewed over time to detect the final carried object locations based on their motion and location priors. Last, the detected carried object regions are accurately segmented. Compared to existing methods, our approach is not only focusing on detecting carried objects. It takes a step forward and accurately segment them. Our method to carried object segmentation couples local appearance cues with location priors of the detected carried objects to produce accurate segmentation. Experimental evaluation on two datasets demonstrates that both our carried object detection and segmentation methods significantly outperform competing algorithms.