Multiple human upper bodies detection via candidate-region convolutional neural network

Upper body detection on images is a challenging task in practical application scenarios and shares all the difficulties of object detection. This paper focuses on the problems of the multiple upper bodies, including the diversity of appearances, the various object scales, and the frequent occlusions. To address these problems, we divide the upper body detection into two stages to form a Candidate-Region Convolutional Neural Network(CR-CNN). In the upper body candidate generation stage, a deep hierarchical model is proposed. This model is built by a graphical model that contains the appearance model and deformable model. The appearance model is built based on the feature maps in a CNN, and the deformable model is defined by each pair of connected parts to compute the relative spatial information in the graphical model. In the upper body candidate refining stage, the detected bounding boxes serve as the candidate regions and refined in the CR-CNN. Moreover, multiple convolutional features are introduced into the CR-CNN to provide the local information and contextual information. The proposed method is compared with the state of the art on the TV Human Interaction (TVHI) dataset and HollywoodHeads dataset. The experimental results demonstrate the effectiveness of the proposed method.