Anomaly Detection System in Video Surveillance using Deep Learning Techniques
Mauricio Andres Aguilera Jaramillo, Daniel Alejandro Armijos Bustamante · 2020
In the field of video surveillance analysis, anomaly detection is an essential task to en- sure public protection throughout detecting frames which tend to contain unusual events such as robberies, assaults, fights, among others. Recently, Convolutional Neural Network (CNN), a category of deep learning techniques, have shown great progress in computer vision tasks, with applications related to both image and video classification / recognition, and more specifically for anomaly detection tasks. Nonetheless, they fail to handle accu- racy in real-scenarios due to the presence of noise, specific context/situations, variability of how different events are defined, limited data for training purposes, high computa- tional resources required to respond on real-time, among others. This work explores two outstanding video classification CNNs architectures to analyze their structucture, their capability to run in real-time scenarios, and the ability to properly classify video frames into normal and abnormal events. Such abnormal anomalies include abuse, arrest, and as- sault. It also introduces a new CNN achitecture, called Frankensnet, focused on anomaly classification from video frame. Frankensnet aims to take the structural characteristic of the explored CNNs, taking into account the accuracy achieved, caoability of detec- tion in real time and training time. Experiments are performed referring to UCF-Crime dataset. As the preliminary results, this paper provides a point-of-view to select CNN architectures for anomaly identification, considering accuracy as well as training and execution time. Moreover, FrankensNet demonstrated to be suitable for scenarios requiring a high accuracy. Nevertheless, training the architecture takes approximately twice as much time as the explored architectures. Finally, a desktop program is also provided to test the performance of each CNN architecture on real-time scenarios with tasks of anomaly detection.