Study Of Pedestrian Detection Based On Improved Fast Feature Pyramids

Pedestrian detection as a hot issue in the object detection field aims to detect up-right pedestrians in the images which include different background. It has been widely applied in the field of intelligent transportation system, driving assistance system, video surveillance and robot etc. So pedestrian detection is of great business value. Technological difficulty of the pedestrian detection are illumination, shadow’s disturbance in complicated scenes and many uncontrollable factors, such as pedestrian’s posture and clothes, random obscured properties etc. Therefore, how to solve these problems as well as research pedestrian detection’s algorithm which is suitable for all kinds of scenes is very important to generalize and apply the pedestrian detection technology to reality.Fast feature pyramids pedestrian detection algorithm is simply expounded in the article, aggregate channel feature is used to extract features in the detected windows.ACF is improved by our article, including template integral feature and off-line selected ACF feature. Then secondary detector is added to reduce the number of false positive detected windows in the fast feature pyramid pedestrian detection algorithm’s detected result.Aimed at the problem that priori knowledge of pedestrian isn’t considered to be used into ACF, template integral feature, where level and vertical differential templates according to up-right pedestrian’s feature are used to weighted sum pixels in the template, then make difference operator, is proposed. The simulation analysis of the results shows that the template integral feature can effectively improve detection rate of the pedestrian.To the problem that Adaboost classification has high possibility to select confused background features as distinguished feature, the off-line selected ACF feature is proposed to eliminate irrelevant information. During the off-line selected period, Adaboost based on decision-making trees is used to select features, some features and their indexes can be obtained, then choosing features whose occurrencefrequency is high as effective features among them. Indexes of the features are applied to the on-line Adaboost training stage, which eliminates the disturbance from confused background features. The simulation analysis of the results shows that the off-line selected ACF feature can effectively improve detection rate of the pedestrian.To reduce number of the false positive detected windows in the result of fast feature pyramid pedestrian detection, secondary detection algorithm of pedestrian detection based on sparse representation classification is proposed. During the training period, positive and negative sample’s features are extracted to build a overcomplete dictionary. A vector of Sparse representation coefficients of uncertain sample is solved in the detection stage, then summing the sparse coefficients of positive and negative samples separately, the ratio of two sums is regarded as determining criterion. As the result of experiment shows that the amount of false positive detected windows is decreased after added secondary detector, the performance of detection system also get improved.