Automated uniform recognition to enhance video surveillance at correctional services in South Africa

Abstract

Injuries from inmate altercations are common in correctional service facilities. Monitoring incidents manually from video surveillance can be challenging. Computer vision has the potential to assist security personnel in securing facilities. This work compares two methods for recognising occupational uniforms with the aim of improving situational awareness and safety in prisons. The first method uses histograms of hue and saturation (HS) colour-space features and a shallow learning classifier. The second method uses convolutional neural network (CNN) models trained with either hand-engineered or automatically learned features. A training dataset with civilians, South African correctional service and police uniforms was created. The experimental results demonstrate comparable performance from shallow learning algorithms and CNN models. Machine learning algorithms evaluated on the proposed colour features achieved an average balanced performance (F1-score) of 0.85 and inference times range from 0.01 to 4.9 milliseconds.