摘要： 针对目前大多数视频异常检测方案在局部异常检测上的不足，提出一种基于局部时空特征的视频异常检测方案。该方案先提取运动描述符，再量化拆分，对每个特征描述符使用不同标度的时间-空间滤波器，获得各时间-空间区域的平滑估计，为训练和测试视频计算出各区域的局部K最邻近(KNN)距离，根据上述局部KNN距离，得出测试和训练视频的总体分值。对总体分值排名，确定异常。将该方案在公共数据集(UCSD数据集、人群异常UMN数据集、U型转弯数据集)上进行测试，结果表明，该方案的误差率、曲线下面积等性能指标优于现有的视频异常检测算法。
关键词:
视频异常检测,
时空特征,
数据维度,
特征描述符,
K最邻近距离,
分值
Abstract: Aiming at the disadvantages of the local anomaly detection in most video anomaly detection schemes, this paper proposes a video anomaly detection scheme based on local spatio-temporal signatures. Motion descriptors are extracted and quantized into small blocks. Spatio-temporal filters at different scales are applied to obtain smooth estimates at each spatio-temporal location for each feature descriptor. Local K-Nearest Neighbor(KNN) distance for each location is computed for training and test video. These local KNN distances are aggregated to produce a composite score for the test and training video. The composite scores are ranked to determine anomalies. To test the performance of the proposed scheme, this paper applies it to several published datasets, such as UCSD dataset, the UMN dataset of crowd anomalies and the Subway dataset. Results show that the performance of proposed scheme is better than the existing video anomaly detection algorithms.
Key words:
video anomaly detection,
spatio-temporal feature,
data dimension,
feature descriptor,
K-Nearest Neighbor(KNN) distance,
score value
中图分类号: