基于支持向量机的复杂背景下的人体检测

常用的人体检测方法多是基于经验风险最小化原理的传统统计理论，其性能只有在样本趋于无穷大时才有理论上的保证，而在实际应用中，学习样本通常是有限的。针对传统统计理论在人体检测中存在的不足，提出了一种基于统计学习理论——支持向量机(SVM)的人体检测方法，利用彩色空间对背景进行自适应建模提取运动目标，然后使用训练好的SVM进行验证是否是人体。为了简化SVM分类器的设计及提高机器学习的效率，提出了一种星形向量表示法用于抽取目标的特征向量，并且用实验方法得到了这种表示法的最优表示。将SVM与ANN进行比较，并且对不同内积函数的SVM的性能也进行了比较。实验结果表明，SVM的性能要优于ANN，并且采用径向基函数的SVM性能最好。该方法鲁棒性强，正确率高，解决了复杂背景下运动人体实时检测的一些关键问题。

Support Vector Machine Based Human Detection under Complex Background

In the field of computer vision, the research on human detection has a wide application prospect. Prevalent human detection methods usually use traditional statistical theory, which is based on empirical risk minimization. But the minimization of empirical risk over limited training data does not imply good generalization to novel test data. Aiming at the shortcomings of traditional statistical theory used in human detection, a new method based on SVM is presented in this paper. An adaptive background subtraction method combined with color is used to segment the motion objects. Then the trained SVM classifier distinguishes the motion object whether it is a human or not. In order to simplify the design of SVM classifier and improve efficiency of machine learning, a center radiating vector representation is proposed to abstract features of the object. And the optimal representation is obtained by experiments. During the machine learning, a bootstrap method is adopted to reduce the complexity of training SVM. Experiments show that the performance of SVM is better than ANN, and the radial basis function SVM has better performance than other SVMs on human distinguish. This method has strong robustness and high accuracy.