Motion detection with deep learning based optical flow

It is a concept that has important applications in many areas such as motion detection and prediction, security (building / workplace), autonomous vehicles, health, traffic, and military. During motion detection, ambient disturbances, light changes, rapid movement of the object and noises make motion detection difficult. The performance of the methods and algorithms used in motion detection is extremely important for accurate motion detection. Optical flow approaches, which form the basis of motion detection methods and are based on many algorithms, also have an important place in motion estimation. Knowing the deep learning-based optical flow calculation and motion detection methods used in motion detection and dating back to 2015 is extremely important in terms of comparing performance and accuracy with current methods. Deep learning based optical flow computation and motion estimation methods and datasets used for optical flow are the main subject of this study. The aim of the study is to examine the studies that have achieved success by using deep learning methods in optical flow calculation, to compare some algorithms with classical methods by retraining, and to perform performance analyzes on different video images. In this study, a literature review was conducted on optical flow calculation methods, which form the basis of motion detection, from traditional approaches to today's deep learning-based approaches. The traditional approaches, Lucas-Kanade, Horn-Schunck, Farneback methods, were applied to different motion pictures and the results obtained were compared. Then, the deep learning-based approaches FlowNet, FlowNet2 and RAFT methods were trained on Sintel, KITTI datasets and the endpoint error and loss values were compared using different hyperparameters through graphs, tables, and visual outputs. In the study, unsupervised learning, and optical flow calculation approaches, which are popular recently, were compared with supervised learning approaches.