Dynamic Visual Words Extraction And Assembly Classification Research For Pedestrian Detection

Pedestrian Detection System (PDS) aims to detect and localize pedestrians in the front view of a moving vehicle through the sensors installed in the car. It also tracks the detected objects and alerts the driver to possible collision ahead of time, so as to make driver-assistance and improve the traffic safety. PDS is one of the core techniques of Intelligent Transportation Systems (ITS), which draws much attention by industry and has great real application value. On the other hand, it is one of the most challenging problems in computer vision due to the appearance variation of pedestrians and the complexity of background. It is a hot cross research topic relates to machine learning, computational intelligence, automation and control. information fusion. Therefore, the study of PDS has great theoretical research value.Compared with general object detection systems, PDS is a dynamic detection system which must meet the demanding of real time application. In the detection procedure, not only the pedestrians have appearance variation, the dymamic variation of road conditions also makes the detection difficult. On the other hand. complex algorithm can not be adopted due to the demand of real time application. Therefore, how to design a rapid detection method which could dynamicly adapt to the road condition is a critical issue for pedestrian detection.Existing pedestrian detection methods can be divided into two main categories, image processing and machine learning. Thereinto, the image processing based methods usually have complex processing which can not meet the requirement of real time detection. On the contray, the classification based methods are the mainstream in this field because of its simplicity and efficiency. In general, a classification based pedestrian detection system contains two steps. In the first step, physical feature descriptors are extracted from some candidate regions of the image. In the second step, a classifier is trained with the extracted features to predict whether the candidate region contains a person.So far, many approaches have been proposed for the feature extraction and classification in PDS. which could get good performance in fixed scene. However, when the road condition varys significantly, the performance of these methods may drop quickly. Fisrtly, the physical feature descriptors are usually sensitive to the variation of road condition, the distribution of feature may vary with the road condition, and thus the extracted features in oringal scene are no longer credible. Secondly, when the road condition changes seriously, the classifier trained in oringinal scene is no longer useful, restraining a classifer using samples in the new scenes could improve the performance, but it is not feasible in real time application. Finally, combining detector trained in different scenes to form an ensemble classifier is a good solution to increase the road condition adaptability, however existing ensemble strategy is proposed for fixed scene, the ensemble algotithm in changing road condition still need further research to get better overall performance.In order to increase the road condition adaptability of existing classification based methods for PDS. this thesis has made corresponding solutions for the three problem mentioned above, containing the design of assembly classifer, the key words selection for pedestrian and the detector for changing scenes. The main work and innovations of this thesis are shown as follows:(1) Considering the drawback of existing ensemble strategy, an evolution based selective ensemble classifier is proposed for pedestrian detection system.Exsiting ensemble algorithms in PDS employ all of their component learners to constitute an ensemble. However, it has been shown that it may be better to ensemble some instead of all of them. In this thesis, we have proposed an evolution based selective ensemble classifier for PDS, in which only an optimal set of the trained learners selected by the algorithm could participate the final voting for the detection. In order to overcome the problems of slow convergence, premature convergence and local optimization, an additional population is employed to cooperate with the host population. Moreover, a time-spending factor is added into the fitness function thus the algorithm finally select the optical combination of classifiers according to the overall performance.(2) Considering that basic physical features are relatively sensitive to the variation of road condition, we extract a key word set of pedestrians.Basic physical features are relatively sensitive to the variation of road condition, so we extract visual words based on the basic features of samples. To improve the representation ability of the visual words, a compact set of key words, which catch typical patterns of pedestrians, are selected to form a standard codebook. For this purpose, a low dimensional model is needed to be built to recover the geometric structure of visual word distribution. In our approach, the visual words are modeled with manifold learning, which not only makes nonlinear dimension reduction, but also establish an internal model of the visual word space. Moreover, a centrality measure based selecting algorithm is proposed to select typical key words on the manifold model. Finally, each person can be represented with some key words of pedestrians.(3) A novel bag of visual words based detector is proposed, which could detect pedestrians in unseen scenes by dynamically updating the key words without retraining.In existing classification method, when the trained detector no longer catches the characteristics of current scene due to the changing of road condition, the performance decreases rapidly. A strategy is to retrain a classifier using the samples caught from the new scenes. However, it is feasible in real time application. In this thesis, a novel approach based on the bag of visual words is proposed to increase the adaptability of the classification based PDS, which is able to adapt to the new scenes without the use of auxiliary data. Instead of retraining the detector, the proposed method trains a fixed detector and adapts the detector to changing road conditions through the adaptive adjustment of the key words in codebook. After the detection procedure of our method, each detected pedestrian sample is further analyzed. Effective new words contain typical pattern are absorbed into current codebook to help detecting pedestrians in the new scenes. On the other hand, the word that no longer reflects the characteristic of the current scene will be excluded.