A study on real-time image recognition SOC with affine motion segmentation

本研究提出了一种新的高精度通用图像识别算法。基于该算法，我们还开发了一个具有实时性能的SOC。SOC由一个通用CPU和一个专门用于仿射运动分割的处理器组成。处理器将图像划分为对应于在立体图像中捕获的对象的区域。处理器计算梯度图像矩以估计左右图像中相同对象的区域之间的视差。CPU采用支持向量机对目标进行分类。支持向量机再次使用图像梯度矩作为图像特征进行识别。然后，运动分割处理器使用下一次的图像，通过仿射运动模型估计来跟踪对象。基于SOC的图像识别可应用于车辆检测、跟踪等领域。SOC的芯片面积为4.1 mm平方。工作频率为94 MHz。处理吞吐量为VGA 91fps。

This study proposes a novel image recognition algorithm for general purpose with high accuracy. We also develop an SOC based on this algorithm with real-time performance. The SOC consists of a general purpose CPU and a processor dedicated to affine motion segmentation. The processor divides an image into regions corresponding to objects captured in stereo images. The processor calculates gradient image moment to estimate a parallax between regions of an identical object in the left and right images. The CPU classifies objects with SVM (Support Vector Machine). The SVM uses the image gradient moment again as an image feature for recognition. Then the motion segmentation processor tracks objects with affine motion model estimation using an image of the next time. Image recogniton by the SOC is applicable to many applications such as vehicle detection and tracking. The chip area of the SOC is 4.1 mm square. The operating frequency is 94 MHz. The processing throughput is VGA 91 fps.