Research And Implementation Of Real-time Pedestrian Detection In Mobile Scene

With the continuous development of the field of computer vision,especially the breakthrough progress deep learning achieved in recent years,Mankind is gradually entering the artificial intelligent age.Pedestrian detection technology has an important application in the robot,security monitoring,unmanned driving and many other artifical intelligent areas.At present,the existing pedestrian detection algorithm has two shortcomings.First,the detection speed is slow,despite the high detection accuracy it achieved.Second,the existing algorithm is mostly based on scientific research and the operating platform is high-performance computer,directly apply these algorithm to the actual scene will face many problems.This thesis study on how to accelerate the speed of pedestrian detection,we present a method to reduce the detection area by using the object proposals and the method of using the disparity to exclude the non-pedestrian window in advance.Morever,This thesis study various methods to reduce computational complexity and spatial complexity for mobile scene and the embedded platform.The specific contents of this thesis are as follows:Firstly,the pedestrian in the image often account for only a small part of the whole image.Most of the areas detected in the actual detection process are the non-pedestrian area.We use object proposals to obtain the detection area and detect pedestrian on it.The area to be detected by using our method is smaller than the area directly detected from the original image.Secondly,In the process of using the classifier to determine the window,most of the windows identified are non-pedestrian windows.According to this feature,We use a method to exclude non-pedestrian window in advance by using disparity.Our method greatly reducing the number of windows need to be identified and speed up the classification stage.Thirdly,in order to detect pedestrian in the mobile scene,where the computing platform has the disadvantages of small memory and low performance,we study a variety of methods to reduce the computational complexity and spatial complexity.In order to reduce the spatial complexity,we optimize current method to reduce the amount of data.In order to speed up the multi-scale detection process,We use a multi-scale and multi-window detection method based on remapping.Sliding on the image with only one scale window can obtained the results of multiple scale windows by remmapping the classifier.In order to eliminate the phenomenon of jitter detection result when detect pedestrian in the video,an optimization method is used to stabilize the detection result.Forthly,we achieve the purpose of real-time detection of pedestrian on the DM8168 embedded platform by making full use of the platform resources and reducing the computational complexity of the algorithm.