基于特征融合的行人检测技术

首先研究了三种不同的特征算子在基于图像行人检测中的应用. 他们分别是: 梯度直方图(Histogram of Oriented Gradient, HOG)、局部三值模式特征(Local Ternary Patterns, LTP)以及改进了的局部三值模式特征(Sqrt Local Ternary Patterns, S-LTP). 对以上三种特征算子进行了实验比较, 最后将HOG和S-LTP算子融合得到HOG+S-LTP的基于多特征的行人检测算子, 利用SVM分离器在INRIA人体库上进行了实验, 实验表明, 融合后的特征显著地提高了行人检测率, 同时也满足实时性要求.

Pedestrian Detection Technology Based on Features Fusion

First of all, we studied three different types of feature operators in the application of image-based pedestrian detection. They are Histogram of Oriented Gradient(HOG), Local Ternary Patterns(LTP) and the Sqrt Local Ternary Patterns(S-LTP). Through conducting experiments to compare these three feature operators, we combined the algorithms of HOG and S-LTP and proposed a new human detection feature operator named HOG+S-LTP. Finally, by using a linear SVM as the classifier, we compared the new feature with the other single features in INRIA person dataset. Results of experiments showed that the new combined feature operator significantly improves the pedestrian detection rate, and it also meets real-time requirements.