Human Posture Classification Using a Dual Deep Convolutional NN With Silhouette Images

This paper proposes a dual deep convolutional neural network (DDCNN) model with two-view images to classify four human body postures, including standing, bending, sitting, and lying. The proposed image-based classification system consists of three parts. The first part segments a moving human body from the images captured from left and right cameras. A minimum enclosing rectangle (MER) of each body silhouette is then found. In the second part, the two silhouette images from the MERs are sent to the two convolutional networks in the DDCNN with AlexNet as the backbone network. The third part concatenates the flattened features extracted from the convolutional networks and performs classification using a fully connected network. Experimental results and comparisons with different deep learning models show the advantage of the proposed classification system.