Fast Pedestrain Detection Based On Hybrid Feature

Face and human detection is a popular computer vision research topic.It has been widely used in various applications,such as virtual reality,intelligent transportation and smart TV.However,it is still difficult to utilize object detection algorithms in real-time system so far because of the slow object detection speed.Computing image feature pyramid has been a common approach in pedestrian detection for improving detection accuracy.However,building feature pyramid is a time consuming task.To accelerate detection speed,we approximate the nearby scale classifier instead of extracting features multiple times form the resizing images.These approximated classifiers can be applied to achieve object detection without image resizing.The main research contents of this work include:(1)A new feature,BPG(Binary Pattern of Gradient),has been introduced in this work.BPG feature is a transformed formation of HOG,and it keeps the characteristic of HOG feature.In order to divide the gradient orientation into 8 bin range and sum up the corresponding gradient values in different gradient directions bin,the mean value of each orientation bin is used as the threshold value.By comparing with the threshold value,the BPG features are obtained by the corresponding two value coding.The experimental result demonstrates that the new feature has a strong pedestrian distinguishing ability.(2)The feature pool consists of four features that are selected based on a lot of experiments.They are BPG feature,LBP feature,gradient values feature and down orientation feature.We analyzed the complementarity of these features.Experimental results show that combination of the foure features achieves 97%detection accuracy.(3)The classifier employed in this work is developed based on the Adaboost algorithm.In our classifier,each stage has diferent recognition abilities.It makes obvious negative examples to be excluded in former stage classifiers and detects hard negative examples in later stage classifiers.The proposed classifier not only improves the accuracy of the classifier,but also reduces the number of detection windows(4)We present a new pedestrian detection method in which several classifiers are approximated in order to replace the image feature pyramid building process.The proposed approach accelerates object detection speed by converting the building process of dense image pyramid to approximate several scaling classifiers.Approximating classifiers may cause slight detection accuracy loss about 1%-3%,but the detection speed can be improved by about 200%.