Pedestrian Detection Technology Research Based On The Random Forests

In recent years, informatization and intelligentization have become a development trend in the field of video monitoring. As an important part of video monitoring system, pedestrian detection has a broad application prospect in intelligent transportation, search and rescue, smart home, etc. Nowadays, many research institutes are engaged in research of pedestrian detection technology, and some targets have been achieved. However, because of the complex and volatile environment around people, with variety and occlusion of people gestures, pedestrian detection now has the problem of low identification rate and high rate of false drop. So it is necessary to keep up research in pedestrian detection. As an important statistical learning algorithm, random forests have a very good performance in classification and regression. It is not sensitive to data noise and outliers, have a high rate of accuracy, and is not easy to be over fitting. Therefore, random forest is used for research of pedestrian detection in this paper. The main work is as follows:Firstly, this paper will discuss the two aspects of the machine learning pedestrian detection algorithm respectively:pedestrian characteristics and classifier. For pedestrian characteristics, the relatively popular feature currently such as Harr, LBP and HOG feature, and the way in which it depict and describe mechanism of information in figure are detailedly discussed. For classifier, the most commonly used detection algorithm such as SVM, Adaboost are studied, and their detection principle, advantages and disadvantages are analyzed.Secondly, random forest algorithm are researched, and based on that, some improvement have been made. Random forest is made up of several independent and identically distributed decision trees. The algorithm uses voting method in classification, but the classification abilities of different decision trees are different. Therefore, some improvement has been made in this paper. Double training method can be used to adjust the weights of classifiers, which can increase the weights of classifiers that have good classification effect, and decrease the ones of classifiers that have bad classification effect. Test on different data sets shows that compared to traditional random forest algorithm, recognition rate and false drop rate of modified random forest algorithm proposed in this paper are improved in some degree.At the last, modified random forest algorithm is used to train the pedestrian detection classifier, and comparison of the classification effect of modified algorithm and that of the traditional algorithm and other algorithms is made. Detection result shows that compared to other algorithms, the classification effect of modified algorithm is improved in a certain degree in pedestrian detection. Then, the modified algorithm is applied in picture and video pedestrian detection system and the research target is achieved.