Color spatial feature-based approach for multiple-vehicle tracking

We present a robust approach to multiple-vehicle tracking, which combines deterministic and probabilistic methods. The observation model is built with an improved color correlogram, which is a feature vector with a compact correlogram using overlapping fragmentation to make the ideal form to measure similarity with a Bhattacharyya coefficient. The observation and state model of multiple vehicles is given under a CamShift framework. To overcome the disadvantage of particle impoverishment, a layered particle filter architecture embedding Camshift is proposed, which considers both the concentration and the diversity of the particles, and the particle set can better represent the posterior probability density. We also present experiments using a real video sequence to verify the proposed method.