Complex Interactive Activity Recognition Based On Spatio-temporal Reasoning

Activity recognition, especially complex interactive activity recognition is very important for modern intelligent surveillance systems, and it has a wide range of commercial applications. In recent years, the video surveillance system has been widely arranged in various places, but now most of the intelligent surveillance systems need staffs in24hours, the technology of activity recognition is urgently needed, this will help save a lot of manpower and resources, but more importantly is that it can detect the behavior in the scene quickly, so it is helpful for precautions and post-surveys. Based on the number of participants and whether there is an interaction behavior in the activity,the activities can be divided into single-target none interaction behavior, multi-targets interactions and group behaviors. In this paper we study the multi-targets interaction activities recognition, we propose an new activities recognition method based on the characteristics of temporal and spatial relations.Target detection is an essential step in activity recognition, and it provides an important prerequisite for the behavior analysis. In this paper first we introduces several methods of moving target detection:optical flow, frame difference method, background subtraction method. First,extracting moving targets in the video, and then extracting the behavioral characteristics, behavioral characteristics is the input of the behavioral analysis, is helpful for behavior recognition accuracy rate. This paper describes several common characteristics and characteristics of representation. Based on spatial reasoning knowledge according to the relationship in time, according to the temporal and spatial variation between moving objects spatial-temporal characteristics of the middle layer is obtained.The article describes the behavior recognition methods commonly used in several other categories:template matching method, the state space method, discriminated law. Hidden Markov Model is widely used in behavior recognition, and gets good results. It is one of the most important methods for behavior recognition. This paper describes the hidden Markov model, including the mathematical background of the model and its three main issues and resolution algorithm. It also presents a few Hidden Markov extensions Model.An interactive activity with multi-objects is usually complex, in this paper, we propose an approach of based on spatial-temporal relation with new extend Hidden Markov Model for interactive activities recognition.Multi-objects activities are naturally hierarchical with spatial-temporal relationship, we propose an approach to extract three different levels features(group,double,single),and propose a new model——Multi-observations Three-Layers Hidden Markov Model (MTHMM) and the method of parameters evaluation and the method of calculating the likelihood. The study shows that the new method of extract features and new model have a good performance in interactive activity recognition with fair good robustness.