Abstract:

In view of the problem of degradation issues as well as the distribution of target class weights adapted to the phenomenon that may arise in the training process of the traditional Adaboost algorithm, the authors introduced a few improved methods to these problems. The article presented a modified Adaboost algorithm based on the adjusted weighted error distribution to limit the expansion weights. In addition, the Adaboost algorithm improved the classifier output forms, i.e., using output of the probability value instead of the discrete value and increased the detection rate more dramatically. Experiment shows that the test rate of the improved Adaboost algorithm could achieve excellent results in the Inria data set. There are good prospects of application in the field of video security surveillance.

Key words: error distribution, Adaboost algorithm, weight update, positive and negative error ratio, classifier output