Abstract: A method is proposed for behavior abnormality detection and normal behavior classification of human in video sequence.In this method，the walk activity and the run activity are defined as normal behaviors，and the main innovative point is the selection of the spatio-temporal characters.Firstly，the region character is selected，so the statistical rule can be acquired by the normal behavior analysis.The behavior which can not accord with the rule is thought of abnormal.Then the outline character is selected too.These outline can describe the spatio-temporal distributing of the target.A behavior is classified into two categories（walk or run） by the training based on Support Vector Machine（SVM）.This method is experimented with some samples.The results show that this method can detect the abnormal behavior and classify the normal behavior.The efficiency is better than the other related works too.

Key words: abnormality detection, behavior classification, Support Vector Machine（SVM）, character selection

摘要： 提出了一种固定场景视频序列异常检测和正常行为分类的方法。该方法定义行人正常的走路和跑动为正常行为，最大的特点在于时空联合特征的选择。首先选用区域特征，通过分析正常行为找到特征的在时间上的统计规律，视频序列中行人不符合规律的行为被判定为异常。然后选用具有时空联合分布特点的目标轮廓特征，通过支持向量机（Support Vector Machine，SVM）进行训练，在训练的基础上判断目标行为是走路还是跑动。该方法在一定样本基础上进行了实验，实验结果表明，该方法能够较好进行行为检测和分类，性能比其他方法也有提高。

关键词: 奇异性检测, 行为分类, 支持向量机, 特征选取