An improved adaboost learning scheme using LDA features for object recognition

Trained detectors are the most popular algorithms for the detection of vehicles or pedestrians in video sequences. To speed up the processing time the trained stages build a cascade of classifiers. Thereby the classifiers become more powerful from stage to stage. The most popular classifier for real-time applications is Adaboost applied to rectangular Haar-like features. The processing time of these detectors is short enough for real-time applications running on low cost hardware, but for difficult object classes the performance, especially for the later stages, drops. That is mainly due to the local rectangular features that cannot separate the object samples from the non-object samples, especially in later stages where the positive and negative samples become very similar. This paper deals with a new approach that combines the local weak features to global features, improving the separation capability of Adaboost classifiers significantly.