Research On Multi-label Classification Method For Vehicle Video

Multi-label classification has attracted much attention in recent years due to its wide application in many fields such as image classification and social network data mining,which has become an important research topic in the field of artificial intelligence.Different from the traditional single-label classification task that each instance is associated with only one label,the goal of a multi-label classification task is to output a set of associated labels,that means each instance is associated with multiple labels.This thesis studies the multi-label classification method for vehicle video,including the feature extraction method for vehicle video and its multi-label classification algorithm.The main innovative contributions of this thesis are as follows:1.Provide a feature extraction method for vehicle video.In order to study the problem of vehicle video feature extraction,some suitable and representative features are selected,then a feature extraction method for vehicle video is designed.This method can extract video features comprehensively and accurately,which lays the foundation for multi-label classification of videos.Vehicle video feature extraction method is divided into two steps: video preprocessing and feature extraction.The video preprocessing: the method grays the video to reduce the memory consumption;and performs filtering operations on the video to remove the influence of noise during video acquisition and storage.Video feature extraction: the method extracts edge features of main objects and lane line features of the road.2.Propose a multi-label classification algorithm for vehicle video: DR-ML-KNN(Dimensionality Reduction based ML-KNN).The algorithm studies the multi-label classification of high-dimensional data and aims at reducing the time of finding the k nearest neighbors of each instance in the training set with large scale.The basic framework of this algorithm is the traditional Multi-Label K-Nearest Neighbor(MLKNN).However,considering that if features are extracted at a fixed frame rate,the size of the training set will be large and the feature matrix of each frame of the video image has a high dimension,so that the idea of dimensionality reduction is introduced.Dimensionality reduction is performed on the transposed matrix of the feature training set of the video,which aims at removing highly similar instances and reducing data redundancy.In this thesis,experiments are performed on 5 different Berkeley vehicle video data(BVRD),which shows that the method can effectively improve the accuracy of label prediction and reduce the algorithm running time by about 50%.3.Propose an incremental multi-label classification algorithm for vehicle video: IDR-ML-KNN(Incremental Dimensionality Reduction based ML-KNN).The algorithm studies the memory consumption of high-dimensional feature data,which aims at reducing the storage of features and speeding up the efficient of the algorithm while ensuring comprehensive feature extraction.IDR-ML-KNN is based on the multilabel classification algorithm DR-ML-KNN.Considering that streaming video data will consume more and more memory,the idea of self-adaptive streaming big data learning algorithm based on incremental tangent space alignment is introduced.The algorithm stores new vehicle video data in an incremental manner,which improves efficiency and accuracy.In this thesis,experiments are performed on 5 different Berkeley vehicle video data(BVRD),which shows that the algorithm can improve the accuracy of label prediction and reduce the algorithm running time.