Snatch theft detection in unconstrained surveillance videos using action attribute modelling

In a city with hundreds of cameras and thousands of interactions daily among people, manually identifying crimes like chain and purse snatching is a tedious and challenging task. Snatch thefts are complex actions containing attributes like walking, running etc. which are affected by actor and view variations. To capture the variation in these attributes in diverse scenarios, we propose to model snatch thefts using a Gaussian mixture model (GMM) with a large number of mixtures known as universal attribute model (UAM). However, the number of snatch thefts typically recorded in a surveillance videos is not sufficient enough to train the parameters of the UAM. Hence, we use the large human action datasets like UCF101 and HMDB51 to train the UAM as many of the actions in these datasets share attributes with snatch thefts. Then, a super-vector representation for each snatch theft clip is obtained using maximum aposteriori (MAP) adaptation of the universal attribute model. However, super-vectors are high-dimensional and contain many redundant attributes which do not contribute to snatch thefts. So, we propose to use factor analysis to obtain a low-dimensional representation called action-vector that contains only the relevant attributes. For evaluation, we introduce a video dataset called Snatch 1.0 created from many hours of surveillance footage obtained from different traffic cameras placed in the city of Hyderabad, India. We show that using action-vectors snatch thefts can be better identified than existing state-of-the-art feature representations.