面向边缘端设备的轻量化视频异常事件检测方法  

Lightweight video abnormal event detection method for edge devices

作　　者：李南君 李爽 李拓 邹晓峰 王长红 Li Nanjun;Li Shuang;Li Tuo;Zou Xiaofeng;Wang Changhong(Shandong Yunhai Guochuang Cloud Computing Equipment Industry Innovation Co.,Ltd.,Jinan 250013,China;State Key Laboratory of High-end Server&Storage Technology,Jinan 250013,China;Qilu University of Technology(Shandong Academy of Sciences),Jinan 250353,China)

机构地区：[1]山东云海国创云计算装备产业创新中心有限公司,济南250013 [2]高效能服务器和存储技术国家重点实验室,济南250013 [3]齐鲁工业大学(山东省科学院),济南250353

出　　处：《计算机应用研究》2024年第1期306-313,320,共9页Application Research of Computers

基　　金：山东省自然科学基金资助项目(ZR2023QF050);国家自然科学基金资助项目(62203242)。

摘　　要：现有基于CNN模型的视频异常事件检测方法在精度不断提升的同时,面临架构复杂、参数庞大、训练冗长等问题,致使硬件算力需求高,难以适配无人机等计算资源有限的边缘端设备。为此,提出一种面向边缘端设备的轻量化异常事件检测方法,旨在平衡检测性能与推理延迟。首先,由原始视频序列提取梯度立方体与光流立方体作为事件表观与运动特征表示;其次,设计改进的小规模PCANet获取梯度立方体对应的高层次分块直方图特征;再次,根据每个局部分块的直方图特征分布情况计算表观异常得分,同时基于内部像素光流幅值累加计算运动异常得分;最后,依据表观与运动异常得分的加权融合值判别异常分块,实现表观与运动异常事件联合检测与定位。在公开数据集UCSD的Ped1与Ped2子集上进行实验验证,该方法的帧层面AUC分别达到86.7%与94.9%,领先大多数对比方法,且参数量明显降低。实验结果表明,该方法在低算力需求下,可以实现较高的异常检测稳定性和准确率,能够有效兼顾检测精度与计算资源,因此适用于低功耗边缘端设备。Existing CNN-based video anomaly detection methods improve the accuracy continuously,which are faced with issues such as complex architecture,large parameters and lengthy training.Therefore,the hardware computing power requirements of them are high,which makes it difficult to adapt to edge devices with limited computing resources like UAVs.To this end,this paper proposed a lightweight abnormal event detection method for edge devices.Firstly,the method extracted gradient cuboids and optical flow cuboids from video sequence as appearance and motion feature representation.Secondly,the method designed a modified PCANet network to obtain high-level block-wise histogram features of gradient cuboids.Then,the method calculated the appearance anomaly score of each block based on histogram feature distribution,and calculated the motion ano-maly score based on the accumulation of optical flow amplitudes of internal pixels.Finally,the method fused the appearance and motion anomaly scores to identify anomalous blocks,achieving appearance and motion abnormal events detection and localization simultaneously.The frame-level AUC of proposed method reached 86.7%on UCSD Ped1 dataset and 94.9%on UCSD Ped2 dataset,which were superior to other methods and the parameters were much smaller.Experimental results show that the method achieves better anomaly detection performance under low computational power requirements,making the ba-lance between detection precision and computing resources,which is suitable for low-power edge devices.

关 键 词：智能视频监控 边缘端设备 异常事件检测 主成分分析网络 分块直方图特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]