The Study, Based On The Adaboost Algorithm Of Pedestrian Detection

Intelligent video surveillance system which owns many good characters has been widely researched and used in many fields in recent years. Pedestrians are the important objects for intelligent video surveillance system. In this thesis the pedestrian detection based on ensemble learning is studied in detail. The main contributions are as follows:1. The effects of different weight updating approaches of Adaboost on the performance of classifiers are analyzed respectively. Thus, an extended weight updating approach is presented, which modifies the traditional weight updating process and limits the false positive rate while keeping the whole error stated. On this basis, another new weight updating approach called adaptive threshold weight updating approach is presented. The new method could limit the false positive rate or false negative rate to a certain threshold and meet the special needs of the strong classifier in a cascade classifier.2. Two classes of features are given: one class is called triangle features which are much more suitable to describe the gesture of pedestrians and the other class is called composite features which are much more suitable to describe the inner area of the pedestrians. A corresponding algorithm to compute the feature value of triangle features and a new way to calculate the rectangle features are represented. Experiments show that: It takes fewer classifiers to compose a single strong classifier through training classifiers with new features and the gained strong classifier could get higher detection rate.3. A video sequences database is created, which consists of about 80 video sequences shot from several real scenes. This database can provide effective experiment data to test the algorithms of object detection, tracking, trajectory analysis, etc, and lay a good foundation of further researches.4. A real-time updating pedestrian detection system is developed. A unique characteristic of the algorithm is its ability to train special cascade classifier dynamically for each individual scene. The adaptive threshold weight updating approach and three classes of new features (triangle features, composite features, rectangle features calculated in a new way) are applied in the training process of a single strong classifier. The system produces fast and robust detection results as demonstrated by extensive experiments performed using video sequences under different environments.