Pedestrian detection in digital pictures

Object detection is one of the fundamental subjects in computer vision studies. It requires collaboration of many image processing and computational intelligence algorithms. In this study, pedestrian detection, one of the most challenging tasks in object detection, is investigated. In pedestrian detection applications, intelligent vehicles detect pedestrians at their path and warn drivers or take precautions to prevent collisions. A variety of sensors or cameras can be used for this purpose. In these systems, a camera is attached to the front panel of the vehicle, or two cameras can be used to have a stereo vision. Mobility of the vehicles makes it difficult to make use of the motion information taken from consecutive images since the video taken from these cameras is tilting continuously. As a solution to this problem, the detection of pedestrians in still images taken from these cameras is proposed. Thus, in this study a pedestrian detection system that uses still images taken from a single camera is investigated. As a result of this thesis, a general object detection system has been created which can be used in detection of other objects as well as pedestrians. Besides, it is shown that the problem of constructing powerful classifiers by using simple ones is an optimization problem which can be solved by using a matrice that we call as error matrice. The algebraical solution of this problem is investigated and it is also shown that Adaboost proposes a practical solution but does not guarantee the best one. This solution is improved by using confidence scores which show the confidence of each classification made by the simple classifiers as well as the result of the classification. A novel feature extraction method is proposed which uses image entropy and it is shown that detection accuracy is improved dramatically when it is used with other feature detection methods. A method which uses helper images that we call as counter images is devised in order to compute the entropic features rapidly. The success of the detection system is considerably increased by the help of these contributions.