Binary Artificial Immune Algorithm for Adaptive Visual Detection

A visual model plays an important role in developing an efficient and robust visual tracker. The visual cues employed in the state-of-the-art models are usually predefined and fixed for all the tested videos. However, the discriminative ability of features usually varies among videos. Therefore, using a fixed set is both redundant and noisy: only a subset of any fixed set will present distinct profiles for modeling. Thus, selecting a highly discriminative cue subset in visual modeling should improve the tracking accuracy. In this paper, an optimization method based on a binary artificial immune algorithm is proposed that selects an effective, discriminative feature subset that is adaptive to specific videos. Specifically, a metric is defined to measure the discriminative abilities of visual models. Then, the visual modeling problem is transformed into an optimization scheme, and a binary artificial immune algorithm is introduced and specially designed to solve the modeling problem. Moreover, to preserve the subset of visual cues that are most adaptive to the tracking condition, the selected cues are assigned adaptive weights and modeled in a Sequential Monte Carlo framework. To show its effectiveness, the proposed algorithm is tested on ten representative videos. The experimental results demonstrate the improvement in tracking performance, the improved tracker performs better or comparable with previous excellent trackers from the literature.