Robust visual tracking using adaptive local appearance model for smart transportation

Smart transportation plays an important role in building smart cities. We can obtain mass data from multi-source and use it to manage transportation in an intelligent way. Images and videos can be easily obtained from various sensors in modern road system. They offer abundant information about the transportation. Therefore, visual analysis is a key point in smart transportation management. In this paper we propose a robust visual object tracking algorithm using adaptive local appearance model, which can be applied to transportation system. As the main challenge of tracking is to adapt to the target's appearance change, we build the model with a local patch dictionary which is composed of a static part and an online updated part. The updating scheme is important to determine the quality of tracking results. We propose a coefficient quality based on sparse representation as the sign of updating and introduce incremental learning to compute the new information to update the dictionary. This strategy adapts the templates to appearance change and helps reduce the drifting problem. Experimental results on several challenging benchmark image sequences demonstrate the proposed tracking algorithm achieves favorable performance when the target undergoes large occlusion, illumination change and scale variation.