Pedestrian Detection Algorithm Based On Adaboost And Significant Information

As a specific problem of object detection, pedestrian detection is related to manysubjects, such as image processing, computer vision, machine learning. It is widelyapplied to many areas, such as intelligent transportation, video surveillance, imagecompressing, and multi-media retrieval. Because of these widely applications,pedestrian detection has attracted more and more attentions from researchers andcompanies and made an extraordinary progress in recent years. However, pedestriandetection faces two difficulties to achieve accurate and real time detection. Firstly,pedestrian itself is non-rigid, it is variety in standing angles (front, side, and back),clothes, and occlusions. Secondly, diversity of external conditions (such as shootingangles and attribute, lighting angle and strength, surrounding objects) brings a bigchallenge to precise detection of pedestrian.In this thesis, based on the method of AdaBoost classifier, we make someimprovements that are described as follows to refine pedestrian detection results:(1) Based on a mixture pool of features and AdaBoost ClassifierThis method composes a pool of mixture features including Haar-like rectangles,improved edge oriented histograms and gradient oriented histograms. First we extractthe mixture features from training images. Then we use AdaBoost algorithm to selectsome of critical features. Finally, we get an accurate and real time pedestrian detectionclassifier.(2) Introduce saliency information to improve detection resultsIn an image, regions where pedestrians appear are often consistence with thesaliency regions, so we introduce saliency information to refine detection accuracy. Weemploy two methods to combine saliency information and AdaBoost classifier. The firstmethod is similar with the salient object detection, we directly use saliency features totrain a pedestrian classifier. In the second method, we generate a saliency map of theimage, and use the saliency degree of the corresponding regions to adjust decisionthresholds of the classifier dynamically. These experiments show that saliencyinformation is an effective tool for distinguishing pedestrians from surrounding environment.Overall, we build a pool of mixture features to describe pedestrians both in pixeland gradient level, use AdaBoost to classify these feature data, and get a fast andaccurate pedestrian detection system. Furthermore, by introducing saliency informationof pedestrians, we improve the detection results without sacrificing the running speed.