基于改进颜色自相似特征行人检测方法.docVIP

基于改进颜色自相似特征行人检测方法 　　摘 要: 　　 　　近年来多尺度行人检测在计算机视觉领域受到广泛关注。传统方法需对图像缩放，在不同尺度计算特征，大大降低了行人检测的速度。颜色自相似特征(CSSF)被提出以克服此不足。针对颜色自相似度特征具有维度高和分类器训练时间长等问题,提出一种改进的颜色自相似度特征。改进的颜色自相似度特征结合行人结构相似度，首先定义了固定尺寸的窗口，然后在不同的颜色空间滑动固定大小的窗口进行特征提取，最后结合自适应增强(AdaBoost)算法构建行人检测分类器。实验结果显示：相对于传统颜色自相似度特征的千万级维度，新的特征只有几千维，特征提取速度和分类器训练速度显著提高，检测效果略有下降；与梯度方向直方图特征(HOG)相比，特征提取速度提高5倍，检测效果基本不变，新的方法在实时行人检测和监控系统中有很好的应用价值。 　　?ス丶?词:结构相似；颜色自相似度特征；多尺度行人检测；尺度不变性；AdaBoost 　　 　　?ブ型挤掷嗪?: TP391.4 　　 　　文献标志码:A 　　英文标题?? 　　 　　Pedestrian detection based on improved color self??similarity feature 　　?び⑽淖髡呙? 　　 　　 　　GU Huijian???┆?*??, CHEN Junzhou 　　?び⑽牡刂?( 　　 　　 　　School of Information Science and Technology, Southwest Jiaotong University, Chengdu Sichuan 610031, China 　　 　　英文摘要)?? 　　Abstract: 　　 　　 　　In recent years, multiscale pedestrian detection received extensive attentions in the field of computer vision. In traditional methods, the input image must be resized with different scales to compute the features, which significantly reduces the detection speed. Color Self??Similarity Feature (CSSF) was presented to overcome this problem. An improved CSSF with lower dimension was proposed for the CSSF whose dimension is too high and time??consuming in the training process of the classifiers. Combined with pedestrian structural similarity, a fixed??size window was defined at first, and then the improved CSSF was extracted by sliding the fixed??size window in different color space. Finally, the pedestrian detection classifier was constructed by combining with AdaBoost algorithm. Test shows that compared with the traditional CSSF whose dimension is ten millions, new feature dimension is only a few thousand, and it can be extracted and trained faster, but detection effect decreases slightly; compared with the Histogram of Oriented Gradient (HOG), feature extraction speed improves 5 times, detection effect is essentially the same. The new method has a good application value in real??time pedestrian detection