Online Motion Agreement Tracking

We present a new appearance model for online multi-target tracking, where a patch-based representation of each target to be tracked is introduced along with a sequential update scheme, which we call “motion agreement tracking” (MAT). This method updates the appearance model of each target online by indirectly evaluating the motion consistency among its local patches. We show that a distance measure based on appropriately reweighted local patches will successfully reduce tracker errors especially those lead to track fragmentation and track switching. We also demonstrate that the performance results of a well-designed online 2D tracker, like the MAT algorithm, can actually measure up to those of state-of-theart offline algorithms on various popular pedestrian tracking benchmarks. Our competitive results are particularly appealing since our technique is very efficient. They also suggest that the role of a proper appearance model may be more important than researchers deemed for video-based tracking, where the majority of previous studies focused on motion dynamics. In essence, we designed a person-specific appearance model with a collection of local image patches by dividing the bounding box of a detected person into a grid representation. Each local patch k is described by a 64-bin color histogram in HSV space. Each patch is associated with a weight wk, which is set to be uniform when the tracker is initialized. We would like to assign a high weight to a stable patch that does not change significantly over time, and a low weight that the patch belongs to the background or represents a fast-changing part of the object. When a detection is assigned to the tracker after solving the trackermeasurement assignment problem, a filtered estimate of the object’s global motion vector v is computed by a Kalman filter. At the same time, the method estimates the local displacement vk of each patch based on a similarity measure. Here, we compare two popular measures in our system: the maximum normalized cross-correlation and the minimum histogram intersection distance. Given the local motion estimates, our method re-weights each patch by checking the agreement between vk and the global motion v. The intuition is that if vk is similar to v, then this local patch moves along with the pedestrian, so it is more likely to be a stable region that does not undergo significant appearance change. By focusing our effort on the most stable patches, we can construct a similarity measure that can distinguish between interacting objects. The discrete level g of the agreement is computed by our implementation as follows: