Random Forests for Object Detection

摘要

We present a method for object detection based on random forests. It is accomplished through the generalized Hough transform paradigm, where object centers are voted from small and local parts in images. Some previous works such as the well-known implicit shape model (ISM) take unsupervised clustering method during the training stage and it often lead to false-positive detections due to random constellations of parts. Thus, we employ a random forest to leam a more discriminative model. We use the KAZE local features to construct a random forest classifier and all leaf nodes in each tree constitute a discriminative codebook model. The codebook model is used to estimate object locations via probabilistic Hough voting. Furthermore, before the test stage, we also adopt a salient region detection step to reduce false-positive detections. Experiments show that our method provides good detection results in complicated environments.