A Deep Learning Approach for Motion Segmentation Using An Optical Flow Technique

There is a severe need of a constant human surveillance of the real-time security footage. Computer Vision is a novel way to reduce human involvement for the said task. Motion segmentation is a crucial step in analyzing video data. The challenges present in motion segmentation, such as illumination changes, dynamic background, and camouflage negatively affect the performance of existing motion segmentation algorithms. In this paper a method of using Convolutional Neural network with optical flow is proposed to improve performance and segment required motion properly. The proposed method is compared with the Lucas-Kanade optical flow method in terms of F1 score. The dataset used is wallflower video dataset. This contains different challenges of motion segmentation viz., illumination changes, dynamic background and clutter.