Object tracking using Mutiple Hypothesis

摘要

The goal of our paper is to develop a algorithm that observes moving objects in a scene and learns observations to study the patterns of activity. The proposed algorithm is based on the Multiple Hypothesis Tracking (MHT) approach. In the proposed approach we first detect the object and track by optimizing trajectories through multiple hypothesis analysis. Here, we focus on tracking motions of objects and learn patterns of activity in a scene. Further segmentation is carried out based on an adaptive background subtraction method that models each pixel of a frame as a mixture of Gaussians and this model is updated using on-line approximation method. The results given by Gaussian distributions are then determined to check which are most probably from a background process. The tracker proposed is stable, which reliably deals with lighting changes, clutter and long-term changes in the scene. By using the hierarchical binary tree classification technique, the image sequences are classified. The tree formed based on the joint co-occurrences found within the sequences. The hypotheses formed are placed according to their distributions. The proposed method is used to track individual instances in a scene as well as sequences.