The Research On Classification Technology In Pedestrian Detection System Of ITS

Pedestrian Detection System (PDS) in Intelligent Transportations Systems is what detects the pedestrians in the view of the vehicle, and takes actions after comprehending the behavior of pedestrians to aid the driving or even drive the vehicle automatically. PDS technology, as the basilica component of the Intelligent Transportations Systems, is an across research topic related to sensor system, hardware system, and machine learning system, etc.Aiming to finding a more effective classification method, the thesis studied the feature extracting and classification technology, analyzed the disadvantage of each technology, summarized the state of the art, and proposes some thought to improve it. The traditional classifiers and its combinations encounter different limitation. Though used widely in pattern recognition, SVM classifier is not utilized adequately in pedestrian detection as its high computational cost.Feature selection performs an important role in pattern classification. It aims at excluding the irrelevant features as more as possible, and obtaining a distinguished subset of the features that can simplify the classifier. The less informative features not only make the learning algorithm weaker, but also conceal the intrinsic characteristic behind the dataset. Along with the development of newly technology, the dataset is becoming larger and larger. As a result, many irrelevant features usually appear in the dataset, which make the traditional learning algorithm encounter to the huge challenge, especially in efficiency and generalization of the classifier. Thus, an algorithm that can remove irrelevant features and weaken the redundancy becomes very essential. A feature selection method based boosting is proposed in this thesis. The method selects subset of features for from the original set by Sequential Forward Selection, using the evaluation function on the kernel space. Experiment show that the method perform better than non-boosting strategy. Compared to the other evaluation function, it can bring more optimal features.As a common method, Multi-classifier combination falls across challenges no matter that in series or in parallel, or even be mixture of both. Tree-like classifier combination is deemed to that combines the advantages of both. The thesis optimized the training process of the tree-like classifier combination, including the selection of negative samples and splitting of the positive samples before training a sub-node of the classification tree, which makes good effects on the classifier. The experiments demonstrated that the non-balance SVM classifier combination is effective in filtering the negative instance accurately.