Human Carrying Baggage Classification Using Transfer Learning on CNN with Direction Attribute

Human carrying baggage classification is one of the important stages in identifying the owner of unattended baggage for a vision-based intelligent surveillance system. In this paper, an approach to classifying human carrying baggage region on surveillance video is proposed. The proposed approach utilized transfer learning strategy under convolution neural network with human pose direction attribute. For this purpose, we first constructed convolution neural network with the target including the presence of baggage and viewing direction of the human region. The network kernels are then fine-tuned to learning a new task in verifying whether the human carrying baggage or not. Rather than using the entire human region as input to the network, we divided the region into several sub-regions and assign them as a channel of the input layer. In the experiment, the standard public dataset is re-annotated with direction information of human pose to evaluate the effectiveness of the proposed approach.