| 摘要: |
| 提出了一种基于运动分析的行为检测方法，用于行人异常行为的检测。利用HSV色彩空间变换法抑制阴影，利用三帧法建立初始化背景模型。将所提取的连续三帧图像的背景像素进行填充融合以实现背景图像的重构，进而更新背景图像，最终完成背景图像的建立。将大津法（Otsu）和背景差分法进行融合以自适应检测前景目标，将目标区域的质心差值、矩形宽高比和倾斜角度的多个特征进行融合，判定异常目标的异常行为。采用国际视频以及自己拍摄的视频进行实验，结果表明该方法能够准确判别行人的行走、跌倒、奔跑行为，并对可能发生的异常行为进行预警，检测准确率最高可以达到97%。 |
| 关键词: 异常行为检测 背景建模 目标检测 特征融合 |
| DOI： |
|
| 基金项目:江苏省自然科学基金项目（BK20201043）；江苏省研究生实践创新计划项目(SJCX20_0701)；南京工程学院产学研专项（CXY201930）；企业委托产学研合作横向项目（科18 092） |
|
| A pedestrian abnormal behavior detection based on motion analysis |
| QIN Binxin,LU Hong,QIU Chun,WAN Wenming |
| (School of Mechanical Engineering,Nanjing Institute of Technology,Nanjing 211167,China) |
| Abstract: |
| A behavior detection method based on motion analysis is proposed to detect abnormal pedestrian behavior.The HSV color space transformation method is used to suppress shadows.The three-frame method is used to establish an initial background model.The extracted background pixels of three consecutive frames of images are filled and merged to reconstruct and update the background image,and finally establish the background image.Otsu method and background difference method are fused to adaptively detect foreground targets.The centroid difference,rectangular aspect ratio and tilt angle features of the target area are fused to determine the abnormal behavior of the abnormal target.International videos and videos taken by ourselves are used for experiments.The experimental results show that the proposed method can accurately distinguish pedestrians walking,falling,and running behaviors,and provide early warning of possible abnormal behaviors.The accuracy of detection is up to 97%. |
| Key words: abnormal behavior detection background modeling target detection feature fusion |
