Learning-based leaf occlusion detection in surveillance video

Leaf occlusion is a challenging issue in surveillance video particularly in public outdoor places. The objective of leaf occlusion detection is to automatically determine whether the video suffers from leaf occlusion or not. To tackle this challenge, this paper proposes a learning-based leaf occlusion detection approach, which incorporates a new feature map into a deep learning network. The proposed approach consists of two key components. First, a spatial-temporal feature map is proposed by exploiting both the hue information in the color domain and the motion direction information in the temporal domain. Second, the proposed feature map is divided into patches and further incorporated into a convolutional neural network (CNN) to develop a learning-based leaf occlusion approach. Experimental results are provided to demonstrate that the proposed approach can effectively detect the leaf occlusion in surveillance video.