Stochastic car tracking with line- and color-based features

Color-based and edge-based trackers have been shown to be robust and versatile for a modest computational cost. However when many distracting features are present it is common for such trackers to get "distracted" and start tracking the wrong object. Using multiple features can reduce this problem - it is unlikely that all will be distracted at the same time. It is also important for the tracker to maintain multiple hypotheses for the state, and sequential Monte Carlo filters (also known as particle filters and used in the well-known CONDENSATION algorithm) have been shown to be a convenient and straightforward means of maintaining multiple hypotheses. In this paper we improve the accuracy and robustness of real-time by combining a color histogram feature with a edge-gradient-based shape feature under a sequential Monte Carlo framework.