基于深度学习特征的异常行为检测

Abnormal Behavior Detection Based on Deep-Learned Features

已有的异常行为检测大多采用人工特征,然而人工特征计算复杂度高且在复杂场景下很难选择和设计一种有效的行为特征.为了解决这一问题,结合堆积去噪编码器和改进的稠密轨迹,提出了一种基于深度学习特征的异常行为检测方法.为了有效地描述行为,利用堆积去噪编码器分别提取行为的外观特征和运动特征,同时为了减少计算复杂度,将特征提取约束在稠密轨迹的空时体积中;采用词包法将特征转化为行为视觉词表示,并利用加权相关性方法进行特征融合以提高特征的分类能力.最后,采用稀疏重建误差判断行为的异常.在公共数据库CAVIAR和BOSS上对该方法进行了验证,并与其它方法进行了对比试验,结果表明了该方法的有效性.

Most existing methods of abnormal behavior detection merely use hand-crafted features to represent behavior,which may be costly. Moreover,choice and design of hand-crafted features can be difficult in the complex scene without prior knowledge. In order to solve this problem, combining the stacked denoising autoencoders （SDAE） and improved dense trajectories, a new approach for abnormal behavior detection was proposed by using deep-learned features. To effectively represent the object behavior, two SDAE were utilized to automatically learn appearance feature and motion feature, respectively,which were constrained in the space-time volume of dense trajectories to reduce the computational complexity. The vi- sion words were also exploited to describe the behavior using the method of bag of words. In order to en- hance the discriminating power of these features, a novel method was adopted for feature fusing by using weighted correlation similarity measurement. The sparse representation was applied to detect abnormal behaviors via sparse reconstruction costs. Experiments results show the effectiveness of the proposed method in comparison with other state-of-the-art methods on the public databases CAVIAR and BOSS for abnormal behavior detection.