Content-aware ranking of video segments

We present an algorithm for ranking videos from multi-camera settings. The algorithm is based on the analysis of objects and events associated to them. Object analysis uses multiple features and is based on motion segmentation using color-based change detection and then trajectories for each target are generated using multi-frame graph matching. Event detection employs a hidden semi-Markov model with duration distribution that generates a sequence of activities performed by each target using Viterbi decoding. For each object size, pose and its associated events (including their duration) are used to rank each frame of a video. The proposed method for ranking is deadline driven. The performance of the proposed approach is demonstrated on standard multi-camera datasets as well as on simulated 3D multi-camera scenarios.