Research On Time Series Classification Algorithm For Interpretable Gait Recognition

Gait recognition is a novel research field in biometric identification.It is widely used in developing security applications due to its non-contact,remote monitoring properties and other advantages.The study of gait recognition has important research significance.One of the important research implications of data mining is to make people understand the key of practical problems,which we call interpretability.In recent years,gait recognition has received extensive academic attention.However,current literatures mostly focus on accuracy,there are few studies discussing interpretability.It is noticed that methods in the field of time series classification can be used for reference to enhance interpretability.of gait recognition.Based on discussion above,this thesis takes interpretability as main consideration,represents the features of gait image sequences as time series.Interpretable time series classification algorithm is improved.While the performance on time series classification tasks is improved,the proposed algorithm also satisfies the requirements of interpreting gait recognition tasks.Further,gait recognition algorithm with both accuracy and interpretability is explored.The main contribution of this thesis includes:(1)Time series classification algorithm based on Shapelet is studied.A Shapelet-oriented random forest algorithm with novel decision tree structure is proposed,classification accuracy and training speed of the random forest are improved.This method embeds two Shapelets from different classes into decision tree node and makes decisions based on distance comparison between the two Shapelets and time series.(2)A forest interpretation method is proposed for the random forest algorithm proposed in the first work.This method can detect time series segments that play an important role in the classification process.Compared to existing methods,the proposed method can provide importance score for different classes separately,thus providing more information that can be understood by human.This feature is the foundation of the interpretable gait recognition algorithm of this thesis.(3)Based on the first and second works,a gait recognition algorithm with both interpretability and accuracy is proposed.This method first uses background subtraction method to extract foreground,then models human body topology in the video,and then converts joint angle,height,centroid coordinates and other features of walking human into multi-dimensional time series.Finally,the proposed pairwise Shapelets forest is extended to the multi-dimensional time series classification problem and applied to the extracted feature time series for gait recognition.Furthermore,in the process of modeling the leg motion,it is necessary to estimate inclination of leg.Existing methods are sensitive to noise.We propose a two-stage regression method that enhances the noise immunity of this step.Experiments show that the proposed algorithm improves the existing interpretable time series classification algorithm.After being extended and applied to the gait recognition problem,it is a method that is both interpretable and accurate.