Acceleration of Probabilistic Occupancy Map-Based People Localization via Down-Sampling of Rectified Images

• Main Authors: Lin, Hui-Mei, 林慧玫
• Other Authors: Chuang, Jen-Hui
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/c52xuf

碩士 === 國立交通大學 === 資訊科學與工程研究所 === 106 === In recent years, vision-based people localization systems have effectively assisted people to deal with huge amount of surveillance data. With the popularity of such systems, improvements in accuracy and efficiency of people localization have got lots of attention. One of the main trends in people localization is based on the probabilistic occupancy map (POM). The POM-based approaches are robust against noisy foregrounds under serious occlusion and adverse illumination conditions, while achieving great localization accuracy. However, such approaches have two main limitations: (i) it is assumed that videos are taken at head or eye level so that rectangular human models and the computationally efficient integral image can be utilized; (ii) to estimate whether there is a person, the POM-based approach will iteratively estimate the probabilities of people locations, which leads to high computation complexity. In this thesis, we propose an innovative image transformation approach. Firstly, line samples extending from the vanishing point of vertical lines (VPVLs) are produced in each view. Then, we conduct horizontal projection for each line sample. The appearance of a person in the transformed (rectified) image will then become upright rather than slanted. Thus, the degradation in localization accuracy due to (i) can be avoided, while the computationally efficient “integral image” can also still be used. Experimental results show that the image transformation proposed in this thesis can indeed improve the accuracy of the POM-based people localization approach under more general surveillance camera configurations. To enhance the efficiency of the POM approach, while retaining the accuracy in people localization, the transformed image is further compressed in horizontal and vertical directions via down-sampling. With less image pixels to be processed, the computation time can be reduced. Experimental results show that the proposed approach can not only accelerate the POM-based approach, but also retain great accuracy of people localization if the compression is performed in the vertical (but not horizontal) direction.