An Ensemble Model For Human Posture Recognition

Abstract

Human Body Pose Estimation (HBPE) and HumanBody Posture Recognition(HBPR) have improved significantly in the past decade. Gaining access to huge amounts of data, Kinect camera, neural networks and specifically deep convolutional neural networks (deep convnets) have led to fascinating success in these fields. In this paper we propose an ensemble model for human body posture recognition. Deep convnets are the main building block and fundamental aspect of our proposed model. We leverage deep convnets in two variations to classify postures. First, we use them for an end-to-end training scenario. We perform transfer learning with Imagenet weights on deep convnets with our gathered dataset of RGB images to classify five different postures. Second, we use a pre-trained deep convnet[1] (pose estimator) for estimating human body joints in RGB images. The pre-trained pose estimator has been trained to calculate a total of 17 2D joints coordinates and we utilize these coordinates to train a decision tree-based classifier for classification among five classes. Both variations are examined with different settings. The best settings for both variations are combined together to create our proposed model. More specifically, the classification layers of both variations are stacked together and fed to a logistic regression unit for a better classification result. Transfer learning, training and experiments in this paper are based on only RGB images from our gathered dataset and human body joints coordinates extracted from these images, which conveys that our proposed model does not require depth images or any sensor. Eventually, experimental results on the images show that the proposed model has higher performance than fundamental variations. Specifically, our model is able to correctly recognize the human posture in the majority of the images that one of the two fundamental variations fails to classify. The code for the proposed model and our gathered dataset are available on github1.