Abstract:
Surveillance cameras are widely used in public areas e.g., streets, shopping malls,
banks etc. to increase public safety. While the monitoring capability has not kept pace
because of the rapid increase in surveillance cameras. So, there should be research
done to address these difficulties. The objective ofthis project is to develop an anomaly
detection system to automatically recognize the anomalous activities from surveillance
cameras. This report explores different techniques used for anomalies detection.
Different stages involving pre-processing stage, segmentation, and feature extraction
will be studied and discussed. Finally, the end product ofthe algorithms will be written
in the software called Jupiter Notebook. This project uses the Deep Neural Network
approach. In our approach, we will have normal and anomalies videos as bags and
video segments as instances in multiple instance learning (MIL), and by using 3D
Convolutional architecture our model automatically learns an anomaly ranking model
which predicts high anomalous scores for anomalous video segments, and we
introduce sparsity and temporal smoothness constraints in the loss function to improve
anomaly during training. The system first proceeds with the pre-processing ofthe data
set by smoothing. Filtering, segmentation, and features extraction are also performed
in the process. Our dataset consists of 1900 long and uncut real-world surveillance
videos, having 13 different anomalies such as fighting, road accidents, etc. as well as
normal videos. It can be useful for two tasks. First, general anomaly detection having
all anomalies activities in one group and normal activities in another group. Second,
for detecting each of 13 anomalous activities. This system is designed for making ease
for the individual user as well as future development.
