Research On Real-time Pedestrian Detection System Based On Convolutional Neural Network And AdaBoost Algorithm

Pedestrian detection mainly studies how to get accurate coordinates of pedestrian targets in images or videos quickly and accurately.Pedestrian detection is widely used and plays an important role in self-driving car,intelligent monitoring,robot vision,human traffic monitoring and human behavior analysis.It takes challenges to detect all pedestrian targets in one image because of the difference between different pedestrian targets.Sometimes even the same person may have a different appearance,which also increases the difficulty of getting the right results.At present,the system with convolutional neural network can achieve better performance than other pedestrian target detection systems.However,the performance is poor when this kind of system detects the small size pedestrian targets in complex environments.This thesis proposes an improved pedestrian detection system to improve the detection accuracy for small targets.This improved pedestrian detection system optimizes the preprocessing process of the data sets,and optimizes the classification network of the detection system with a kind of improved calculation method of feedback loss.And in order to improve the detection accuracy of small size pedestrian targets,a kind of classification subsystem based on HOG feature and a kind of detection subsystem based on SSD framework are designed.The main system is responsible for obtaining the exact coordinates of large size targets and the region proposal of small size targets.And the subsystem calculates the accurate coordinates of the small size pedestrian targets.In order to evaluate the performance of the improved algorithm,this thesis uses Faster R-CNN detection framework and SSD detection framework to construct the main structure of the pedestrian detection system,and uses inception network structure,residual network structure and deep decomposable convolutional network structure to construct the feature extraction network of the main system.In addition,in this thesis,Caltech pedestrian detection dataset and CityPersons pedestrian detection dataset are used to evaluate performance of the system.The experimental results show that the improved pedestrian detection system has high detection performance when running in real time.