Anomalies in a specific video scene are defined as unlikely behaviours that differ from the dominant behaviours in that scene, highly dependent on human workers, most current monitoring systems fail to detect anomalies over time, moreover, Analysis based on anomalies detected in videos helps the state to improve public safety management.
Today a large number of surveillance cameras are connected directly to each other,therefore an automatic system is necessary for analyzing the videos.
The objective of this work is to understand how to detect events in order to create a desktop application that can
detect events in video surveillance.
Our approach divided into three main steps, first we
divide the video into a set of sequences then images to simplify data extraction, then discover people from an input image or video. specific, finally to have detected the events we have using convolution neural networks, then frames each clip of the video according to the emergency (red if the current event and predicts a violent situation, else the frame will be green).
We use the sequential model, it is suitable for a simple stacking of layers where each has exactly one input tensor and one output tensor, starting from 32 to 64 with a 'relu' type activation function, for the size of the filters we choose (3,3), this layer will be followed with a maxpoolin2D, which shortens the input representation samples by taking the maximum value with pool_size for each dimension.
Top comments (0)