Abnormal Behavior Detection Based on Deep-Learned Features

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 vision words were also exploited to describe the behavior using the method of bag of words. In order to enhance 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. © 2017, Editorial Department of Journal of Hunan University. All right reserved.