Concept drift adaptation in video surveillance: a systematic review

The world we live in is dynamic by nature. Frequently, the environment changes in ways we cannot predict. In machine learning, the phenomenon that occurs when a model has its prediction effectiveness degraded due to unforeseen changes is known as concept drift. Applications of smart video surveillance tend to suffer from concept drift due to changes in illumination, weather, and scene structure. This work differs from previous ones as it brings focus to the problem of concept drift from a surveillance video perspective which presents additional challenges compared to other sources of data, such as high dimensionality, spatial and temporal relations between data, and real-time constraints. The approaches and algorithms used to cope with concept drift are compared and discussed. We also present datasets and metrics used to evaluate the effectiveness of the algorithms.As contributions, we present a new classification of concept drift adaptation methods, delineate the characteristics and limitations of techniques that deal with concept drift, and analyze practical aspects, such as real-time processing and memory constraints. Moreover, we conclude that informed concept drift adaptation methods have been employed 90% less than continuous adaptation ones.Research directions include using established concept drift detection techniques applied to surveillance video data, exploring datasets for concept drift in surveillance, strategies to deal with the high dimensionality and volume of surveillance video data when adapting existing models, and the creation of frameworks to manage drift adaptation while applying computer vision tasks.