Pedestrian Detection Based On Integral Channel Features

Pedestrian detection has been widely used in surveillance,human-computer interaction,driverless,robotics,and Advanced Driver Assistance Systems.It is one of the most research areas of computer vision.High accuracy and speed are goals pursued by researchers,but it is difficult to consider both of them at the same time.Pedestrian detection based on the integral channel features can achieve better accuracy,and at the same time,it ensures the detection speed.Many real-time systems use this pedestrian detection method.The research on this detection method has not yet reached saturation.Thus there is room for further improvement in detection accuracy and this is also the goal of this thesis.The main research contents of this work include:(1)When training classifiers with the AdaBoost algorithm,it is critical to build a suitable feature pool.This thesis proposes a "selected features" method.First we find features with good classification results by studying features with different sizes one by one and then use these selected features to build the feature pool.The advantage of this method is that the feature pool is small,it requires less computer memory and short training time.More importantly,a better classifier can be obtained by using this method.(2)The sum of pixels within a rectangle is usually used as the feature value.This thesis proposes a method called "weighted summation" to compute the feature value.The purpose is to increase the weight of some region in rectangle in order to enhance the richness of the feature.The experimental results show that this method improves the detection quality.(3)A multi-classifier detector is designed and implemented to further improve the detection performance which reduce the miss rate by 2~3 percent points.