基于多尺度融合和双记忆单元的视频异常检测OA

In the task of weakly supervised video anomaly detection,existing methods often neglect the learning of the temporal patterns of normal behaviors due to excessive focus on the modeling of abnormal features,resulting in a blurred boundary for distinguishing normal from abnormal in the feature space,which in turn leads to problems of positioning deviation and high false alarm rate.We propose a video anomaly detection model based on multi-scale fusion and dual memory units.This model designs a multi-scale fusion module(MSTF)to extract the details of short-sequence actions and the context of long-sequence events in parallel.After adaptive convolution fusion,a feature representation containing multi-granularity temporal semantics is generated to solve the problem of context loss in traditional single-scale convolution.The dual-branch dynamic position attention mechanism(DPAM)is proposed.The time branch is adopted to model the inter-frame dependency pattern.The content branch uses sine and cosine position coding and weighted fusion of at-tention scores to accurately capture the temporal dependency between video frames and significantly improve the temporal positioning accuracy of abnormal events.The introduction of channel gating mechanisms reduces information overlap and noise interference between different channels,and strengthens abnormally correlated channels,effectively solving the problem of information redundancy during feature fusion.The experimental results show that the AP index value of the proposed method on the XD-Violence dataset reaches 82.42%,and the AUC index value on the UCF-Crime dataset reaches 86.71%,which is superior to most mainstream methods,especially the performance on the XD-Violence dataset is more outstanding.It is 2.8 percentage points higher than the baseline and has certain competitiveness.