A Fast and Robust Embedded Optical Flow Unit on Motion Estimation

Mubashara Rehman, Chung-Ming Own, Ziqi Li

Abstract: In the last few decades, deep learning is already being widely used to yield excellent performance, automate processes, pattern detection and natural language processing. Among different types of deep neural networks, convolutional neural network has been most extensively studied. In this study, the authors have introduced a network architecture named Fast Unit of Optical Flow (FUOF), which enables the network to estimate optical flow through a fast and robust approach. Our proposed method, FUOF based on the updating of FlowNetC, which includes a layer that correlates feature vectors at different image locations. We also introduced the Extended Sobel method to extract the spatial information to optimize further computation time. Moreover, to fuse the spatial and temporal information flexibly and comprehensively, two different fuse networks are defined as FUOF-Sum and FUOF-Concat in our study. The fuse mode's main difference is on the summarizing of each channel for concatenating of channels for the feature extraction this simple yet powerful idea validated by the experiments with MPI Sintel and Flying Chairs datasets. According to our results, the network with FUOF can achieve a competitive efficiency on the accuracy of standard error measure, convergence speed, and robustness for optical flow estimation.