Low operational amount silhouette extraction and its application to video understanding based on statistical processing

该研究项目是建立一个只有10%的操作量的视线/视觉功能，相比传统的，为了给机器人以及移动的终端的视线能力，因为这些应用程序应该由电池供电。视线/视觉的基本功能是基于提取的轮廓的人在场景中。这个基本功能由前景对象的清晰分离组成，然后从前景对象中消除阴影;这两个功能在传统方法中都需要大量的操作。为了达到目标的性能，新的修改多尺度Reinex和一个新的阴影假设已被采用。最后将得到的轮廓应用于步行识别试验，证明了该方法的有效性。为了显示电池支持应用的可能性，开发了前景分割处理器：FPGA功耗为150 mW，相当于ASIC设计的30 mW。

The research project is to establish an only 10% operation amount sight/vision function, compared to the conventional one, in order to give the sight capability to robots as well as mobile terminal, because these applications should be powered by battery cells. Thesight/vision fundamental function is based on the extraction of silhouette of person in a scene. This fundamental function is composed of the clear separation of foreground objects, followed by the elimination of shadows from the foreground objects; both of the functions require tremendous operations in the conventional approach. In order to achieve the target performance, new modification of Multi-scale Reinex and a novel shadow assumption have been employed. The resulted silhouette has also been applied to walk identification trial and shows the effectiveness. In order to show the possibility of battery support applications, foreground segmentation processor has been developed: the result of power dissipation is 150mW for FPGA, equivalently 30mW for ASIC design.