Research On Activity Recognition Based On Markov Logic Network And System Implementation

With the rapid development of sensor and information technology, wearable devices, such as mobile phone, sports wristbands and smart watches are getting more and more popular. Activity recognition have brought people many creative applications, especially in smart home, medical monitoring and e-commerce fields. So activity recognition is not only a meaningful practical problem with wide application, but also a challenging research problem, attracting interests from researchers.The aim of activity recognition is to track and recognize human normal activity from data sets collected by computers, sensors, cameras and other devices. In artificial intelligence, logic-based approach and statistical relational learning are being used widely in activity recognition.This paper focuses on log-based activity recognition. Unlike traditional activity recognition methods, we use a new statistical relational learning method, Markov Logic Network(MLN). MLN combines logic and probability with their strengths: logic rules are good at describing background knowledge, while probability is good at handling fuzzy, uncertainty and noise data.Basing on Markov Logic Network, we propose an activity recognition method which recognizes human activity from log data. The method is divided into three steps: the first step is to use first-order logic formulas to describe background knowledge about specified areas. The background knowledge is either from human or discovered by our Apriori algorithm which mining strong association rules from training data. Second step is to train the network with training dataset, learning weight for each rule. Thirdly, apply this network to do activity recognition.With this approach, we design and implement a system in recognizing student’s online study activity. The system records, stores students’ behavior data, recognizes study activity from the collected data and visualizes the inferred results. We make experiment on the activity recognition method with the real collected sensor data and student behavior log data. The experimental result verifies the feasibility and effectiveness of our method.