Reconhecimento de atividades humanas violentas em videovigilância utilizando redes neurais profundas e delimitação de área de interesse.

Abstract

The growth in the number of video surveillance cameras deployed for monitoring environ- ments in recent years is not proportional to the human capacity to analyze the captured scenes.

Captured scenes may contain evidence of crime occurrences. However, video surveillance

cameras are rarely used to stop or predict criminal activities simultaneously with their occur- rences. To make crime fighting more efficient, the recognition of human actions could be

performed automatically by means of computational techniques capable of detecting and clas- sifying the types of human behavior. Moreover, in the scenario of pattern recognition in video

surveillance systems, another major challenge is to define a threshold between violent and non-violent events in changing environments and behaviors with ambiguous interpretations, considering the context in which they are performed. For this reason, as the nature of the scenes captured from video surveillance cameras consists mostly of common or non-violent behaviors, scene monitoring requires precision and accuracy on the ability to analyze and perceive aggressive actions. In this work, a proposal for the detection of violent human behavior through computer vision techniques is presented, having as main contribution the delimitation of the area of interest of the frame through the Gaussian filter, as well as the reduction of the space of input features for the model, keeping the most relevant features.

Furthermore, the proposal is able to reduce the use of VRAM (Video Random Access Me- mory) by approximately up to 45% during the training phase. The proposed approach obtained

accuracy of 86.5% in the test phase with the RWF-2000 dataset and outperformed the baseline approach, consisting of a convolutional neural network (CNN) trained for the classification of violent human scenes, combined with the technique of cutting the area of interest from the video frames. The approach also outperformed other state-of-the-art proposals in the video surveillance scenario. A performed analysis, pointed statistical significance when adopting the method proposed in this research. The proposal was also evaluated on benchmark datasets in human fight scenarios.