基于颜色自相似度特征的实时行人检测

行人检测在智能监控和辅助驾驶等方面有广泛的应用。当前行人检测中主流特征是梯度方向直方图(HOG),但其计算耗时导致检测速度慢。该文提出了一种新的颜色自相似度特征(CSSF),在颜色通道上计算两个选定的矩形块的比值衡量自相似性。首先,CSSF在描述行人的结构信息的同时避免了耗时的方向梯度计算,具有速度快的优点。其次,CSSF是标量特征,能高效快速与AdaBoost级联分类器结合学习行人检测器。再次,CSSF具有尺度不变性,能快速地进行多尺度检测。针对CSSF含有的海量特征,该文提出增量AdaBoost算法有效学习CSSF特征。实验结果表明:基于CSSF的行人检测器检测精度优于传统的HOG检测器,速度提高了7倍,在640×480大小的图像上达到实时效果。

Color self-similarity feature based real-time pedestrian detection

Pedestrian detection has wide applications in intelligent surveillance and driver assistant systems.The most commonly used feature in pedestrian detection algorithms is a histogram of the oriented gradient(HOG),which is computationally intensive and results in slow detection speed.This analysis uses a color self-similarity feature(CSSF) that calculates the ratio of two rectangles to measure the self-similarity on the color channels.First,when extracting the human structure information,CSSF avoids the time-consuming gradient calculation which increases the speed.Secondly,CSSF uses a scalar feature which can be efficiently integrated with the AdaBoost based cascaded classifiers learning framework for training the human detector.Thirdly,CSSF is scale invariant,resulting in fast multi-scale detection.Tests show that the CSSF based detector gives improved accuracy and 7 times speedup compared with HOG detectors and can achieve real-time processing with 640×480 images.