Real-time one-dimensional motion estimation and its application in computer vision

Abstract

Traditional optical flow methods are designed to generate two-dimensional (2D) motion fields. However in some applications, only the motion field along a predefined line is needed, while the rest of the redundant 2D region (Cong et al., CVPR, 2009) can be ignored. In this paper, we propose a general framework for estimating one-dimensional (1D) motion field using the L 1 norm metric, which can be flexible for adding any constraints depending on the specific intention. Compared with 2D methods in model calculation, which often obtain a local optimal solution, our 1D version has the advantage of real-time global optimization and sub-pixel accuracy with a large displacement capacity. The synthesized experiments have justified the effectiveness of our 1D model. Two vision applications are investigated: (i) for crowd counting, we design a novel unified framework for both line of interest and region of interest counting using our 1D motion model to mosaic the dynamic blobs. Such a framework can adapt to new site without scene-specific learning and overcome overfitting. (ii) For visual odometry, we apply our 1D model to estimate the linear velocity of the robot in real time and obtain a comparable result with the benchmark Encoder. Experiments verified that our proposed model can produce satisfactory results in real time.