Context-aware activity recognition by Markov logic networks of trained weights

Abstract

One of the general objectives of visual surveillance is to recognise abnormal activities from images. Current object detection/tracking techniques cannot directly classify such activities as fighting and snatching, while they reliably recognise primitive actions, such as walking and running. We represent each target activity as ground, weighted and undirected trees, Markov logic networks (MLNs), starting with primitive actions at the bottom and activities on top, using Horn clauses. The likelihood of one ground activity at root gives a reliable probability that the event actually happens. Computing such a probability could be intractable unless the truth values of all the nodes in a given network are known in advance. This study proposes two methods to infer such unknown values in exploitative and explorative manners. An additional modification of MLNs is also considered to improve accuracy of recognition. The experiments by means of unknown value inference methods and modification of MLNs present that these approaches overcome several well-known limitations that the conventional researches have experienced.