Locally connected graph for visual tracking

Based on fast feature extraction, the subspace representation model provides a compact notion of the “thing” being tracked rather than treating the target as a sparse feature representation. The main challenges of the subspace representation model can be attributed to the difficulty of handling the appearance variability of a target object. In this paper, we present a subspace learning algorithm based on graph embedding that uses a Locally Connected Graph (LCG). By constructing a supervised graph with several types of labeled target samples, the algorithm can effectively learn the semantic subspace modeling for some appearance variability. Moreover, by using an additional constraint connection among several subgraphs, the algorithm can obtain a more compact subspace model. In comprehensive experiments, our algorithm has demonstrated better performance than alternatives reported in the recent literature.