Robust Person Detection by Classifier Cubes and Local Verification ∗

Classifier grids have shown to be an alternative to sliding window approaches for object detection from static cameras. However, existing approaches neglected two essential points: (a) temporal information is not used and (b) a standard non-maxima suppression is applied as post-processing step. Thus, the contribution of this paper is twofold. First, we introduce classifier cubes, which exploit the available temporal information within a classifier grid by adapting the local detection likelihood based on preceded detections. Second, we introduce a more sophisticated post-processing step to verify detection hypotheses by comparing a local figure/ground segmentation to a provided prototype model. Experiments on publicly available data demonstrate that both extensions improve the detection performance.