In recent times, there has been a rise in the amount of disruptive and offensive activities that have been happening. Due to this, security has been given principal significance. Public places like shopping centres, avenues, banks, etc are increasingly being equipped with CCTVs to guarantee the security of individuals. Subsequently, this inconvenience is making a need to computerize this system with high accuracy. Since constant observation of these surveillance cameras by humans is a near-impossible task. It requires workforces and their constant attention to judge if the captured activities are anomalous or suspicious.
As a result, this drawback is creating a need to automate this process with high accuracy. Moreover, there is a need to display which frame and which parts of the recording contain the uncommon activity which helps the quicker judgment of that unordinary action being unusual or suspicious.
Therefore, to reduce the wastage of time and labour, we are utilizing deep learning algorithms and autoencoders for anomaly detection. Anomalies are defined as events that deviate from the standard, happen rarely, and don’t follow the rest of the “pattern”. According to Towards Data Science’s September 2019 publication. Depending on your exact use case and application, anomalies only typically occur 0.001-1% of the time — that’s an incredibly small fraction of the time. Autoencoders are a type of unsupervised neural network consisting of an encoder and decoder. It accepts input images and reconstructs them in a latent space. The reconstructed images are then fed once more to the autoencoder and it returns an error margin. The larger the error the more likely the input is an anomaly/outlier.
The goal is to automatically identify signs of aggression and violence in real-time, which filters out irregularities from normal patterns. We intend to utilize different Deep Learning models (CNN and RNN) to identify and classify levels of high movement in the frame. From there, we can raise a detection alert for the situation of a threat, indicating the suspicious activities at an instance of time.
In a control room monitoring environment, the software executes the autoencoder training script over the first 48 -72 hours on each video stream. During this time the system learns in an unsupervised fashion what is the norm on each stream, it considers object shape, object size, as well as travel direction & speed.
Maintaining a pleasant user experience for the operator, this means very few false alarms presented, fast processing time in determining anomalies and therefore presenting video clips to the operator for their attention or intervention once anomalies have been detected. The continuous improvement of the AI model and user experience takes place by the system interpreting operator behaviour, such as acknowledged and dismissed events.
The video analysis and processing of anomalies takes place at the edge and sends only anomaly events to the cloud and control room operator. This approach is bandwidth savvy, whilst maintaining high performance accuracy.
PAISA Technology VMS with anomaly detection is fast becoming the preferred (computationally) lightweight solution for many control room operators and security officials.