基于改进变分自编码器的异常事件检测算法

Abnormal Behavior Detection Algorithm Based on Improved Variational Autoencoder

为了更好地解决复杂监控场景异常检测问题,构建了一种基于改进的变分自编码器的算法基础框架。第一阶段,将原视频帧序列作为输入,通过使用卷积神经网络作为编码器网络,可以更加高效地提取密集分布于视频帧中感受野的外观和运动特征。第二阶段,与基于重建或生成的解决方法相比,该算法假设所有正常事件都服从一个高斯混合模型,而异常事件无法被高斯混合模型下的任何一个高斯分量所拟合。最后,在常用异常行为事件检测的数据集上对该算法进行了测试评估,并与多种方法进行了比较,结果显示所建立的算法快速准确地检测出了异常行为事件并优于其他方法。该算法在异常行为事件的定位方面具有一定的优势。

To address the challenge of detecting anomalies in complex surveillance scenarios preferably,an improved variational autoencoder-based algorithm framework was constructed.In the first phase,the original video frame sequences are used as input,and a convolutional neural network is employed as the encoder network to efficiently extract appearance and motion features from densely distributed receptive fields within the frames.In the second phase,in contrast to reconstruction or generation-based approaches,this method assumes that all normal events adhere to a Gaussian mixture model,while abnormal events cannot be fitted by any Gaussian component within this mixture model.Subsequently,the algorithm is evaluated on commonly used datasets for abnormal behavior event detection and compared to several other methods.The results demonstrate that the algorithm rapidly and accurately detectes abnormal behavior events,outperforming other approaches.Therefore,this method exhibits a significant advantage in terms of localizing abnormal behavior events in complex surveillance scenarios.