Semantic Motion Segmentation Using Optical Flow and Convolutional Neural Networks (CNNs)

N. Gururaj

The literature for motion segmentation of image sequences is fairly large and growing. The recent works indicate the importance of using image sequences captured using a static or a moving camera which influence the performance of such motion segmentation algorithms. However, motion segmentation can become complex for sequence of images captured using a moving camera. In order to boost the performance of such motion segmentation algorithms semantic labeling of sequence of images along with popular motion detection techniques like optical flow have been used. In this thesis work, motion segmentation for sequence of images captured using a moving camera is performed. For designing this motion segmentation algorithm, a classical computer vision approach in combination with a recent and well known deep learning approach is used. More specifically, optical flow orientations along with convolutional neural networks (CNNs) are used for this task. In order to obtain superior motion segmentations for moving camera images, the thesis work uses semantic labeling for sequence of images and optical flow orientations which are calculated using this information. Also, a CNN is used for which these optical flow orientations are used as inputs. The performance of the algorithm will be evaluated for multiple sequence of images.