摘要
针对带有记忆模块的异常行为识别模型存在存储容量有限和有限的时间轴相关性的局限性,提出带有时间和空间注意力机制的异常行为识别模型。该模型在MAND框架下,引用时间和空间注意力机制,通过该注意力机制对卷积层(CNN)输出的特征图从时间和空间两个维度进行加权操作。该方法有效结合时间和空间上下文信息,捕捉局部和全局的特征,输入到解码器中进行重构,然后计算重构帧与输入帧的误差。采用重构当前帧和预测未来帧(下一帧)的异常检测方法,使用相同的网络体系结构分别进行实验。实验结果证明,与其他检测算法相比,所提方法在UCSD Ped2数据集和The CUHK Avenue数据集上的检测精度有所提升,预测方法在两个数据集中的帧级AUC分别提升了4.9%和1.2%,重构方法同样得到了提升,验证了该方法的有效性。
The abnormal behavior recognition model with memory module has the limitation of limited storage capacity and limited time axis correlation.The abnormal behavior recognition model with temporal and spatial attention mechanism is proposed.The time attention mechanism is introduced in the MAND framework,where the feature graph output from convolution layer is weighted from time and space dimensions.This method can capture local and global features by combining temporal and spatial context information,and input them into the decoder for reconstruction,and then calculate the error between the reconstructed frame and the input frame.We use an anomaly detection method of reconstructing the current frame and predicting future frames(the next frame),with the same network architecture for experiments.Experimental results show that compared with other detection algorithms,the proposed method is improved in terms of detection accuracy in the UCSD Ped2 dataset and The CUHK Avenue dataset is improved,and the prediction method is improved by 4.9%and 1.2%in frame-level AUC in the two datasets,respectively,and the reconstruction method is also improved,which verifies the effectiveness of the proposed method.
作者
付孟丹
宣士斌
王婷
李培杰
FU Meng-dan;XUAN Shi-bin;WANG Ting;LI Pei-jie(School of Electronic Information,Guangxi Minzu University,Nanning 530006,China;School of Artificial Intelligence,Guangxi Minzu University,Nanning 530006,China)
出处
《计算机技术与发展》
2023年第8期51-58,共8页
Computer Technology and Development
基金
国家自然科学基金(61866003)。