Research On Computational Classifier Ensemble Model And Application For Pedestrian Detection

Pedestrian Detection Systems(PDS) aim to detect and localize pedestrian in front of a moving vehicle accurately and quickly,with the help of some sensors and intelligent processing algorithms;and then to forecast the pedestrian-vehicle related collision;at last,to alert the driver and to make some emergency control if necessary. This is a hot research topic extracted from industry requirements;it is an important part of Active Safety,Smart Vehicle and Intelligent Transportation Systems,which has great research value and marketing value.At the same time,PDS is also a cross-research topic related to sensor technology,machine learning,automation and control,information fusion and computational intelligence.At present,PDS research can be categorized into two types:1) researchers from auto industry trend to use expensive sensors,such as infrared cameras, millimeter-wave radars and laser scanners,to guarantee higher detection accuracy and speed;2) some other researchers prefer to develop simple and cheap solutions with optical cameras only.They expect to get acceptable detection accuracy and real-time detection speed with algorithm improvement.Some researchers considered that vision based PDS has some irreplaceable advantages and the key detection technologies(e.g. classification) can be easily extended to systems based on other sensors.Therefore, vision based PDS also has high research value and application value.This thesis focuses on optical camera based pedestrian detection systems and key technologies.This type of PDS has following main difficulties:1) the autonomic and irregular movement of the detection platform and objects;2) the various ad time-variant scenes;3) the diversity of human appearance and partly-sheltered problem.Therefore,classification becomes a key technology for PDS which needs to be well developed.At the same time,a practical and low-cost vision based PDS antitype system is also great needed by researchers and auto industry.Classification for PDS has three main difficulties:1) for the pedestrian detection problem,there are too many objects need to be classified in a single frame and most of them are non-pedestrians;2) for classifier training,the samples are imbalanced too; 3) classifiers for PDS need to satisfy the three conflicting demands at the same time, and it is hard to find a balance point of high detection rate,low false positive rate and high detection speed.At present,classifiers adopted in PDS have following shortcomings:1) single classifiers have imbalanced performances(e.g.low detection rate,high false positive rare and low detection speed),and they are not suitable for varying scenes.2) serial ensemble classifiers have low false positive rate and high detection rate;however,their detection rate is comparatively low and still not suitable for varying scenes.3) parallel ensemble classifiers have high detection rate and low false positive rate,and can be applied in varying scenes;however,their detection speed is very low and can not satisfy practical requirements.To solve the difficult problems mentioned above and to conquer the shortcomings of existing technologies,we considered it necessary to purposely design a high-performance classification model and corresponding algorithm at first;and then to study its application methods;at last,to build up a practical on-board PDS antitype system based on the proposed model and algorithm.Around this research topic,several contributions are made in this thesis:1.In order to get a classifier with banlanced and optimal overall performance, tree classifier ensemble model was proposed,which has both advantages of serial classifier ensemble and parallel classifier ensemble.With the single classifier performance model,the quantified expressions of three performance indicators of the tree classifier ensemble can be obtained,and then the classifier design problem turns to an optimization problem.This makes a high performance classifier for practical PDS applications possible.2.The classifier ensemble can get the optimal performance only when proper values are chosen for the key parameters.At present,these parameters are tuned by repeat experiments which cost too much time and still can not get the global optimal values.In this thesis,a computational model was built for the tree classifier ensemble based on the character of sample imbalance.The key parameters can be calculated directly with the computational model,and this greatly accelerates the optimal values searching and guarantees the balanced and optimal performance of the classifier ensemble.Furthermore,a RBF based searching method is proposed to solve the situation when the computational model is not well suitable;this method can also find the global optimal values of the classification model with acceptable time cost.This method shows that the computational model might be used in other backgrounds.3.To satisfy the PDS antitype system requirements of ITS market,two systems are developed based on the computational classifier ensemble model.One is based on PC platform,which is used for offline algorithm verification,and the other one is based on a DSP chip,which is used for on-board pedestrian detection.The PC based system has functions of pedestrian localization,collision forecasting and alarm besides pedestrian detection;furthermore,pedeatrian detection technologies based on dual optical cameras are also tested on this system.The DSP based system can perform real-time pedestrian detection with the help of some technologies such as integral Adaboost algorithm;it is low-cost,power-saving and size compacted,which makes PDS marketization possible.