Body orientation estimation with the ensemble of logistic regression classifiers

Orientation of human body is an important feature that can be used for behavioral analysis in surveillance systems. This cue contains useful information such as the direction of movement or attention. Difficulties such as low quality images, cluttered background and partial occlusion harden orientation estimation. In this paper, we propose a novel approach for determining body orientation using the ensemble of logistic regression classifiers. The logistic regression is a discriminative model that is very efficient in time and space complexities. In addition to these desirable properties, we show that this classifier provides a good classification performance in our problem. These classifiers are trained using Histogram of Oriented Gradient (HOG) descriptors which are extracted from four regions in the bounding box of the subjects. Two types of regions are considered in our method: static and dynamic regions. Static regions include: the whole body, upper half and the lower half of the body. Dynamic region includes the region of head and shoulder, which is located dynamically in various images and should be localized for each instance. To enhance the output of each classifier, we propose a weighting scheme based on the inherent characteristics of the orientation estimation problem and finally combine these outputs in an ensemble method to improve the accuracy. Experimental results show the superiority of the proposed method in accuracy and time complexity as compared to the state-of-the-art methods.