Improved Spatio-temporal Action Localization for Surveillance Videos

We present an improved spatiotemporal action localization framework that operates in an online manner. Current state of the art approaches have achieved remarkable results mainly due to the advancements in action recognition models. These approaches have commonly followed a two-stage pipeline consisting of a region proposal stage and an action classification stage. Recently, the improvement in spatiotemporal action localization models have focused on improving the action classification stage. As a result, the outputs generated in the region proposal stage are suboptimal. We believe that the proposal stage remains a crucial component in determining the overall model performance. As a result, we adopt a tracking model in place of the existing proposal models to generate more accurate and robust regions of interest (RoI). We evaluate our approach on a private CCTV surveillance dataset and on the challenging JHMDB-21 benchmark. We are able to achieve promising results on our private dataset and achieve good results for the JHMDB-21 benchmark.