Pictorial Structural Models, Learning and Recognition in Image Sequences

Abstract

This paper describes the detector of people as assemblies of individual parts as head, torso, limbs, etc. We build on Felzenszwalb and Huttenlocher's approach for efficient assembling of candidate parts into pictorial structures. A human body is represented as a collection of parts arranged in a deformable configuration. The appearance of each part is modeled separately, and the deformable configuration is represented by spring-like connections between pairs of parts. We implemented three different appearance part detectors. First detector is simple, it requires segmented data - it is obtained with background subtraction. Second one uses more flexible color based segmentation of parts. The third one uses SVM classifier for computing probability, important features are extracted from filtered picture as brightness values of pixels. Probability model of appearance of parts and joints between pairs of parts is learned from manually labeled training data. The best match is found as model, which maximizes a posterior probability, or the 100 samples from posterior probability are taken with Monte Carlo method. Then the best match is selected by another method. We propose a new method for selecting the best match from samples based on epipolar geometry, two calibrated cameras are needed. Distance of significant points of the model from corresponding epipolar lines is measured. All the detectors were tested on real video sequences, hundreds of frames long.

Movies

• Tracking.avi
• Two_cameras.avi