Human body part detection using likelihood score computations

Detection and labelling of human body parts in videos or images can provide vital clues in analysis of human behaviour and action. Detecting body parts separately is considerably difficult due to the huge amount of intra-class variations exhibited. In most methods, researchers tend to impose some connectivity or shape constraints on the classifier output to obtain the final detected body parts. In this paper, we propose a novel idea to compute likelihood scores for each of the initial classified body parts based on Bayes theorem using Extreme learning machine´s (ELM) output value (different from the predicted class label). Also, we do not impose any other constraints on the initially detected body parts. We use Histogram of oriented gradients (HOG) features and ELM for initial classification. We also employ a voting scheme that uses inter-frame detected segments to filter out errors and detect body parts in the current frame. Experiments have been conducted to show our method can identify body parts in different body postures quiet appreciably.