Feature selection for real-time tracking

Abstract

We address the problem of selecting features to improve automated video tracking of targets that undergo multiple mutual occlusions. As targets are occluded, different feature subsets and combinations of those features are effective in identifying the target and improving tracking performance. We use Combinatorial Fusion Analysis to develop a metric to dynamically select which subset of features will produce the most accurate tracking. In particular we show that the combination of a pair of features A and B will improve the accuracy only if (a) A and B have relative high performance, and (b) A and B are diverse. We present experimental results to illustrate the performance of the proposed metric.

Comparing discrimination and CFA for selecting tracking features

“…We showed that this diversity could be used to select between rank and score fusions of all three features [10] or, in combination with a relative performance measure [9] to select which pair of three features in addition to rank versus score fusion, for improved tracking accuracy. However, relative performance is only available as a metric in groundtruth experiments, and hence can not be used as part of a general purpose video tracker.…”

“…In our previous work, we have used an averaged, normalized color feature [9] for target tracking. Collins and Liu [2] point out that an average feature is insufficient for background discrimination since it assumes discrimination is a Gaussian problem (see Fig.…”

“…In [10], S consisted of average rank and Mahalanobis score combinations of the three feature measurements: averaged target color, target bounding-box ratio and target region centroid location. In [9], S consisted of the 11 fusion operations consisting of all combinations of 2 and 3 features fused by average rank combination or Mahalanobis score combination of these same three features. Both [9] and [10] use ground-truth experiments to show that a diversity measure calculated from the rank-score functions of each feature can be used to determine which of the fusion operations to apply to generate improved performance.…”

“…In previous work, the authors have developed a multiple hypothesis based [3] multi-target video tracker that employs combinatorial fusion for improved tracking accuracy [8] [9][10] [11]. The input to the tracker is a sequence of video images or frames that the tracker processes one by one.…”

“…The growth of the pool of track hypotheses is exponential. Our approach has been to use combinatorial fusion to address the growth of the hypothesis pool while maintaining tracking accuracy even in cases of multi-target occlusions [9][10].…”