复杂环境中一种基于深度学习的异常检测方法

Anomaly Detection Based on Deep Learning in Complicated Circumstances

为了解决复杂环境中异常检测的问题,提出一种基于深度学习的检测方法。首先,通过引入(you only look once,YOLO)检测,将卷积神经网络回归法提取的物体时空特征,输入到长短期记忆模型(LSTM),追踪复杂环境中个体的运动轨迹。然后,评估相邻个体间运动轨迹的依赖性。最后,采用编码-解码框架训练LSTM模型,预测物体未来的运动轨迹;根据物体未来运动轨迹的异常概率,最终完成异常检测。实验结果表明,解决了复杂环境中运动物体间的相互干扰问题;在时间和空间鲁棒性评估上,优于其他轨迹追踪的方法,从而证明了本方法的有效性和可行性。

To solve the problem of anomaly detection in complicated circumstances,a method based on deep learning has been proposed.First of all,you only look once(YOLO)detection is chose as a regression problem,spatio-temporal features of each object extracted by convolutional neural network(CNN)are fed into long short term memory(LSTM)model,which used to track each individual.Secondly,motion relation of coherent is evaluated between individuals.Finally,LSTM encoder-decoder is adopted to predict the future tracklet of each object,anomaly detection will be done by the abnormal probability of future tracklet.Experiments showed that the interference problem that exists among motional objects is solved by proposed method in complicated circumstances.It is better than other trajectory tracking techniques whatever on temporal or spatial robustness.The proposed method is demonstrated effectively and feasibly.