A Deep Learning Architecture for Recognizing Abnormal Activities of Groups Using Context and Motion Information

Résumé

Currently, the automation of activity recognition of a group of people in closed and open environments is a major problem, especially in video surveillance. It is becoming increasingly important to have computer vision architectures that allow automatic recognition of group activities to make decisions. This paper proposes a computer vision architecture capable of learning and recognizing abnormal group activities using the movements of the group in the scene. It is based on the Activity Description Vector, a descriptor capable of representing the trajectory information of a sequence of images as a collection of local movements that occur in specific regions of the scene. The proposal is based on the evolution of different versions of this descriptor towards the generation of images that will be input of a two-stream classifier capable of robustly classifying abnormal group activities. Moreover, it includes context information to provide extra information to classify the activities including it as the third stream of the classifier resulting in a robust architecture for one class classification problems. The architecture has been evaluated and compared with other approaches using Ped 1 and Ped 2 datasets, obtaining a high performance in abnormal group activity recognition.