Video analytics technology has been developed into wide practical scenarios especially image analysis from surveillance or CCTV camera. The useful information can be extracted from the image such as person detection and tracking, the direction of person’s movement, person counting, human feature extraction such as skin color, clothes’ color, hat wearing, sunglasses wearing, height estimation, male/female/child classification, etc. Another useful application is Image Super Resolution to reconstruct the higher resolution of the face and license plate for better recognition performance.
Face recognition
Face recognition consists of 3 steps: face detection, feature extraction, and feature matching with face with the database. Presently, Convolutional Neural Network (CNN) has been employed to construct model for highly accurate face recognition. The proposed use case can identify person and detect suspicious incidents. In addition, face recognition can also be extended to the application in safety, finding mission persons, health assessment, and retail businesses. The face recognition system can be put into the cloud server where the camera with simple face extraction algorithm can be put in the hardware attached to the camera at the edge of the 5G network and later make processing and decision at the cloud server.
Face recognition device (proposed by CU and Thailand patent pending) that will be used to test with 5G project is in small size. It is easy to install and can work either as a standalone unit or connect with local server to 5G network. The device can recognize up to 8 faces at the same time with the recognition rate higher than 90% with less than 1 second per face on the processing time. In standalone mode, the face database in the device can store up to 4,000 faces.
Anomaly Detection in video scenes
Anomaly detection in video scenes is one of the increasing popular research in computer vision. It can be applied in the specific locations or in crowded scenes to find any possible anomaly, i.e., abnormal behavior or incident, in the scene such as at the entrance/exit of the train station, pedestrian in the public space, people walking in the building, etc. The challenging in anomaly detection is how to extract objects’ feature in spatial and temporal dimension for model recognition. In addition, the system has to be designed well to accurately detect any anomaly in the scene. Other than the accuracy performance, the system needs to be flexible to detect various kinds of anomaly scenes. As in different types of video content, one scene can consist of many anomalies happened in the scene at the same time. We are working toward applying real-time anomaly detection in different of scene at Chulalongkorn University.
The concept of anomaly detection where the highlighted color denotes anomaly objects detected in the scene.