APPARATUS AND METHOD FOR LEARNING TIME SERIES DATA BASED ON MULTILAYERED RANDOM FOREST

摘要

The present invention relates to an apparatus for learning time series data based on a multilayer random forest, and more particularly, to an apparatus for learning time series data based on a multilayer random forest, wherein a candidate region in each frame of training data composed of a time series image of a plurality of consecutive frames a candidate detection unit to detect a tube generator for generating a tube by connecting candidate regions detected in a plurality of consecutive frames of the training data; a sub-tube generating unit configured to generate sub-tube by dividing the tube into a predetermined number of frames to reflect changes over time; a feature extraction unit for extracting features from the sub-tube; and a time-series multi-layer random forest classifier comprising a plurality of layers corresponding to each sub-tube, using the feature extracted from each sub-tube in the feature extraction unit as an input of the layer corresponding to each sub-tube. It is characterized in that it includes a learning unit for learning the time series multi-layered random forest classifier. In addition, the present invention relates to a method for learning time series data based on a multi-layer random forest, and more particularly, to a method for learning time series data based on a multi-layer random forest. detecting a candidate region in each frame; (2) generating a tube by connecting candidate regions detected in a plurality of consecutive frames of the training data; (3) creating sub-tubes by dividing the tube into a predetermined number of frames to reflect changes over time; (4) extracting features from the sub-tube; and (5) a time-series multi-layer random forest classifier comprising a plurality of layers corresponding to each sub-tube. It is characterized in that it comprises the step of learning the time-series multi-layer random forest classifier as an input of. According to the multilayer random forest-based time series data learning apparatus and method proposed in the present invention, a tube is created by connecting candidate regions detected in learning data composed of time series images of a plurality of consecutive frames, and the tube is divided. By extracting features from a sub-tube and performing learning by using the features extracted from each sub-tube as an input for each layer to a time series multilayer random forest classifier comprising a plurality of layers, the learning parameters to be maintained are It provides fast computation speed due to a small and simple decision-making process, and can operate much more flexibly than RNN-based models.