Dense flow field algorithm using binary descriptor and modified energy function

• Published in: 2017 IEEE/SICE International Symposium on System Integration (SII) pp. 1016 - 1021
• Main Authors: Dong-Sung Pae, Hyeon-Chan Oh, Sang-Kyoo Park, Tae-Koo Kang, Myo-Taeg Lim
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2017
• Subjects: Complexity theory; Computational efficiency; Histograms; Image resolution; Machine vision; Mathematical model; Visualization
• ISSN: 2474-2325
• DOI: 10.1109/SII.2017.8279356

In this paper, we describe a Dense Flow-Field algorithm for moving detection of an object using a binary descriptor and a modified energy function. Among the moving detection algorithms, a Dense SIFT-Flow algorithm is recently introduced. In the conventional Dense SIFT-Flow, a SIFT descriptor and an energy function are employed to make the flow vectors containing the movement information of each pixel at entire image. The matching process in the conventional SIFT-Flow algorithm uses descriptor information and a message-passing method in a coarse-to-fine scheme. Although the matching performance of the Dense SIFT-Flow is good for detecting the movement of each pixel, large computational time is needed. To reduce the complexity of the description part, the proposed method employs a binary descriptor. The process of the binary descriptor is simple enough to reduce the complexity. In addition, the energy function in the conventional Dense Flow-Field must be modified for the binary descriptor as replacing the unfair displacement term of the conventional energy function with a fair displacement term. From the experimental results, we can know that the proposed method is faster than the conventional method with respect to making flow field and more robust with respect to diagonal movements.