摘要
提出了一种基于视频人体运动状态判决的的跌倒检测方法,该方法由运动目标检测、目标运动跟踪和目标运动行为识别三部分组成。在运动目标检测方面采用两次目标框选策略提高目标检测精度;利用目标运动轨迹的连续性,具体为利用上一帧运动物体的中心坐标信息来降低目标跟踪的计算复杂度。采用两级支持向量机(SVM)决策的方法实现目标运动行为的识别:第一级SVM分类器利用高宽比等运动物体特征将人体的直立姿态与非直立姿态进行区分;第二级SVM分类器利用Zernike矩特征等特征将人体的跌倒状态从非直立状态中区分出来。初步实验测试表明所提出的跌倒检测算法的性能与光照条件、跌倒方式、摄像头的架设方式均有密切关系,平均正确检测率为88.7%。
A human body falling detection method based on posture recogniton was proposed, which is consisted of three stages： moving target detection, target moving tracking and identification of behavior status. The performance of moving target detection was improved with two level marquee target detection. And continuity of body movement was used to reduce the computation complexity, in which the coordinates of previous flame was utilized. A two-level Support Vector Machine （SVM） classifier was designed to detect body falling, in which human body posture was classified into up-fight or not first with features such as rate of height to width of the moving target; And body falling was selected out from the non-up-right postures with features such as Zernike moments. The experimental results show that the performance correlates closely to lighting condition, the posture of camera and human body falling pattern, and the average accuracy of falling detection is 88.7%.
出处
《计算机应用》
CSCD
北大核心
2014年第A01期223-227,264,共6页
journal of Computer Applications
基金
广东省东莞市高校基金资助项目(201110815600150)
关键词
人体跌倒检测
智能监控
姿势状态
目标定位
支持向量机两级分类器
human body falling detection
intelligent monitoring
posture judgment
object location
two-level SupportVector Machine （SVM） classifier
