Boosted pedestrian detector adaptation in specific scenes

In detector adaptation, the quality and quantity of collected online samples are of fundamental importance, yet have not been thoroughly investigated. In this paper, we present an efficient detector adaptation approach with a novel unsupervised online sample collection scheme, which can obtain sufficient aligned samples in a specific video. Unlike other methods that collect samples by only leveraging the detection confidence or track, we select aligned samples by evaluating the alignment scores using a pixel-wise Gaussian Model. Since this selection would lead to an inadequate number of positive samples, we synthesize positive samples by composing the pedestrian foreground in each aligned positive samples with the scene background at different locations. In this way, we can obtain a large number of qualified aligned positive samples encoding new scene information. With sufficient samples, we adopt a simple yet effective method to obtain an adaptive detector, which not only preserves the effective part of the offline boosted detector but also well adapts to the new scene by adding some new trained classifiers. Experiments demonstrate the efficacy of our sample collection scheme and that our approach significantly improves the performance.