【摘要】 文中针对视频信息中的长距离时间特征关系易被忽略的问题，提出了一种基于全局上下文网络的弱监督视频异常行为检测方法。为了提升对视觉场景的全局理解，提高异常检测的准确性，对时间特征提取模块进行改进，仅计算一个与查询位置无关的全局注意力矩阵，并对所有查询位置共享，有效降低网络计算量和参数量。同时进行网络模块优化，加快运算速度。实验结果表明，基于全局上下文网络的视频异常行为检测算法能够在网络更具轻便性、运算效率更高的情况下有效提高异常检测准确率。更多还原

【Abstract】 A weakly-supervised video anomalous behaviour detection method based on global context network is proposed to address the problem that long-distance temporal feature relationships in video information are easily ignored. In order to enhance the global understanding of the visual scene and improve the accuracy of anomaly detection, the temporal feature extraction module is improved by calculating only one global attention matrix independent of the query location and sharing it for all the query locations, which effectively reduces the amount of network computation and the number of parameters. Meanwhile, network module optimisation is carried out to speed up the calculation speed. The experimental results show that the video anomalous behaviour detection algorithm based on global context network can effectively improve the accuracy of anomaly detection with a lighter network and higher computing efficiency.更多还原