Hierarchy Random Forest And Its Application In Human Activity Recognition

As life goes faster, with the popularization of internet, people have raised higherdemands on applications such as advanced human-machine interaction and virtualreality. Human activity recognition is to detect, recognize, track and finally make ananalysis and description of human activity from a series of videos or images that containhuman body. Human activity recognition has a really extensive application foregroundand a great commercial potential in many new fields and cross-fields, and has becomeone of the hottest research topic in computer vision.Random forest is a new classification and prediction model algorithm. As a goodensemble algorithm, random forest has been a hot topic in computer vision because ofits superior speed, strong resistance to noise, naturally dealing with multi-classproblems, naturally avoiding over-fitting, etc. Random forest has been broadly appliedin many fields and cross-fields, such as biology, economics, medical science, etc.Firstly, a hierarchy random forest algorithm is proposed and applied to humanactivity recognition by improving traditional random forest algorithm. We usespatial-temporal interest points to represent the features of samples, and constructseveral weak classifiers randomly. Then we construct some weak classifiers at thebottom with the features generated by weak classifiers on the top based on thedifferential information of the classes, and ensemble all those weak classifiers into arandom forest, which is called hierarchy random forest. We made detailed analysis onthe algorithm and gave a hypothesis of the algorithm. The analysis and experimentalresults show that the hierarchy random forest algorithm is effective and robust in humanactivity recognition.Secondly, we considered the imbalanced problem in actual application. Bymanually constructing imbalanced dataset, we found that imbalanced problem a ffectsrandom forest greatly. Then, an experiment of hierarchy random forest on ourimbalanced dataset is carried out, which shows the priority of hierarchy random forestas compared with original random forest.