A new method for violence detection in surveillance scenes

Abstract

Violence detection is a hot topic for surveillance systems. However, it has not been studied as much as for action recognition. Existing vision-based methods mainly concentrate on violence detection and make little effort to determine the location of violence. In this paper, we propose a fast and robust framework for detecting and localizing violence in surveillance scenes. For this purpose, a Gaussian Model of Optical Flow (GMOF) is proposed to extract candidate violence regions, which are adaptively modeled as a deviation from the normal behavior of crowd observed in the scene. Violence detection is then performed on each video volume constructed by densely sampling the candidate violence regions. To distinguish violent events from nonviolent events, we also propose a novel descriptor, named as Orientation Histogram of Optical Flow (OHOF), which are fed into a linear SVM for classification. Experimental results on several benchmark datasets have demonstrated the superiority of our proposed method over the state-of-the-arts in terms of both detection accuracy and processing speed, even in crowded scenes.

Discriminative Dictionary Learning With Motion Weber Local Descriptor for Violence Detection

“…This again demonstrates that the proposed approach is significantly superior in performance to all other approaches. The performance of the RVD method [14] and the AMDN [48] on this dataset is consistent with their performance on the Hockey Fight dataset. Furthermore, our MoIWLD combined with our proposed sparse classification method outperforms SRC methods.…”

“…To evaluate the performance of our proposed ideas, we compare our method against the state-of-the-art approaches either implemented by us or cited from literature, including the BoW based methods, the RVD violence detection method in [14], the Appearance and Motion DeepNet (AMDN) method in [48], the Violent Flow (ViF) method in [13], the method in [27] and our recently published method in [24]. To evaluate the classification accuracy, we employ the 5-fold cross validation test on each dataset.…”

“…We compare the results with not only HOG, HOF, HNF, MoSIFT, MoWLD and MoIWLD paired with BoW, but also ViF, RVD and AMDN. The RVD method [14] adopts a Gaussian Model of Optical Flow (GMOF) to extract candidate violence regions, which has reduced many noise disturbances, so its performance is better than the BoW-based approaches. The AMDN approach [48] utilizes optical flow as the input image feature, and there exist many redundant and interference features, so its performance is not very good.…”

“…The dictionary size is again fixed to 1800 in this set of experiments. [14] 85.29±0.16% 0.8878 AMDN [48] 84.22±0.17% 0.8562 SRC [27] 82.7±0.14% 0.8538 MoWLD+Sparse Coding [24] 87.07±0.13% 0.8928…”

“…However, the performance of this method degrades significantly when dealing with crowded scenes. Zhang et al [14] proposed a fast and robust framework (referred to as RVD) for detecting and localizing violence in surveillance scenes, and experimental results on several benchmark datasets have demonstrated the superiority of this method over the state-of-the-arts in terms of both detection accuracy and processing speed.…”