基于多尺度特征记忆增强的异常行为检测算法

Abnormal behavior detection algorithm based on multi-scale feature memory enhancement

针对传统视频异常行为检测任务中存在目标对象空间尺寸变化差异大和对异常行为预测的泛化能力过强等问题,提出一种基于多尺度特征记忆增强的视频异常行为检测改进方法。通过多分支结构的空洞卷积组成的多尺度特征模块在高级特征图上提取不同尺度的特征信息,级联记忆增强模块存储正常行为特征以削弱泛化能力。在多尺度特征模块和记忆增强模块的协同工作下能够有效收集和记忆正常行为场景中的多尺度特征信息。以实验分析验证该方法的有效性。

A video abnormal behavior detection improvement method was proposed to address issues such as significant variations in target object spatial dimensions and excessive generalization capability for abnormal behavior prediction in traditional video abnormal behavior detection tasks.Different scales of feature information were extracted on the high-level feature map by a multi-scale feature module composed of multiple branches with dilated convolutions,and a cascaded memory enhancement module was used to store normal behavior features to weaken the generalization capability.The coordinated operation of the multi-scale feature module and the memory enhancement module was used to effectively collect and memorize the multi-scale feature information in normal behavior scenes.The effectiveness of this method was verified through experimental analysis.