Unsupervised Moving Object Detection with On-line Generalized Hough Transform
Generalized Hough Transform-based methods have been successfully applied to the object detection, especially to the task of pedestrian detection. Such methods rely on the class-specific codebooks to cast probabilistic votes for pedestrian hypotheses. However, systems based this kind of method require the hand-labeled training set to construct the codebook off-line. To overcome the limitations, we propose an unsupervised moving pedestrian algorithm with on-line Generalized Hough Transform. Our contributions are two-fold: (i) an unsupervised instance selection algorithm for automatic training data labeling ; (ii) an on-line Generalized Hough Transform method to for class-specific codebook adaptation. In our method, motion detection together with the proposed instance selection scheme is used to optimally select training samples from the video. The collected samples are then fed into the on-line learning algorithm to construct a set of Extremely Randomized Trees. We verified the performance of the algorithm on two public corpora of surveillance videos.
Keywords: Hough transform, on-line learning, pedestrian detection,