A classification and clustering method for tracking multiple objects

Tracking multiple people in a video poses many challenges due to frequent occlusion and false detection. Classification-based methods have proven to increase the accuracy of multiple object tracking algorithms. However, we propose that instead of training person specific classifiers, we can train video specific classifiers for the classification task. We propose a joint classification method for tracking each object as a class. First, we adopt an offline approach that generates tracklets, classify and cluster them for multi-object tracking using the tracklet affinity framework. Typically, clustering is done after the classification to ensure that objects belong to the same class and are linked temporally. Secondly, to determine the identity of each tracklet cluster, we formulate it as a multi-class classification problem with a Bayesian constraint and solve it using the Gaussian pattern classes algorithm. Finally, we perform experiments using four widely used multi-object tracking sequences. The results of our experiment show that our proposed method outperforms several state-of-the-art multi-object tracking algorithms.