A new pedestrian detection method based on diverse AdaBoost-SVM optimization ensemble classifiers algorithm

作者：Zhang, Yang^*; Liu, Weiming; Deng, Rufeng; Mo, Chen

来源：Journal of Computational Information Systems, 2011, 7(14): 5076-5085.

• Binary tree structure
• Complex scenes
• Component classifiers
• Data sets
• Discriminating power
• Dynamic adjustment
• Ensemble classifiers
• Feature vectors
• Full binary tree structure
• Haar features
• Haar-like features
• Illumination invariant
• Learning abilities
• On-machines
• Parallel connections
• Pedestrian detection
• RBF kernel function
• RBF kernels
• Real time performance
• Robust performance
• Series connection structures
• Urban environments

Under complex scene urban environment, in order to detect pedestrians in images efficiently and accurately, we propose a high real-time and robust performance pedestrian detection method based on machine vision. Firstly, the novel features called Locally Assembled Binary Haar-like (LABH) are selected as the feature vector. In these features, Haar features keep only the ordinal relationship named by binary Haar feature, then, assemble several neighboring binary Haar features to improve the ability of illumination invariant and discriminating power. Furthermore, a new algorithm named Diverse AdaBoost-SVM Optimization Ensemble Classifiers (DASOEC) is proposed. This method used RBFSVM (SVM with the RBF kernel) as component classifiers in AdaBoost. Through dynamic adjustment the values of the RBF kernel function parameters, a set of component classifiers with different learning abilities and moderately accurate is obtained. Then, DASOEC measures the disagreement between classifiers, and selects optimization ensemble classifiers. Finally, a full binary tree structure is presented to build on an efficient classifier structure, which has advantages of both series connection structure and parallel connection structure, and brings in a principle of "Early-rejection", could improve system's real-time performance. The experiment carried out on videos from INRIA, MIT and Daimler datasets illustrates that the proposed method is real-time and feasible enough for pedestrian detection in intelligent vehicle environment.

• 出版日期2011-12
• 单位华南理工大学