Pedestrian Detection Based On Deep Learning

As an important component of object detection,Pedestrian detection has been widely applied in many research fields,such as intelligent surveillance,automatically driving,driver assistance system,intelligent robot and so on.At present,deep-learning based pedestrian detection method has entered a rapid development stage.Compared with traditional pedestrian detection whose feature is man-made,pedestrian detection method which is based on deep learning,using learning human feature extracting through convolutional neural network,can gain a competitive advantage in detection accuracy.But there are still many problems remaining to be solved,such as the problem of missing detection of pedestrians in small size.An urgent problem also arises as to how to achieve a good trade-off between detection accuracy and detection speed.In this paper,the system structure of human detection is studied as well as detailed introduction of feature extraction and classifier training.Based on study and comparisons of various structure of neural networks,we propose several pedestrian detection methods based on deep learning.The main research contents and innovations are as follows:1.A multi-scale pedestrian detection method based on convolutional neural networks is proposed.We analyze some factors and their impacts on the performances of pedestrian detection,these factors include increment of detection layers,parallel convolution layers,changes of the size of convolution kernels and the ratio of size of human in network.The experimental results on the KITTI dataset show that our proposed method can achieves good performances.Parallel convolution layers can add nuanced feature of human,and multiple layers detection can take good advantages of different layers' detection abilities,and it is helpful to do proper adaptations of the ratio of human in the region proposal network.All these adaptions will contribute to the detection of small size human.2.A multitask pedestrian detection system combining traditional method with deep learning method is proposed.The C4 pedestrian detection,whose detection speed is comparatively fast,is ingeniously designed as the first stage of the system,and the detection result is further judged in second stage,which is performed by the third network of MTCNN.Due to the superiority in detection speed of traditional method and high detection accuracy of deep learning method based on big data,our method combining both of them can maximize the detection speed while maintaining a proper accuracy.3.A method combining pedestrian key points detection with pedestrian detection is proposed.The location of key points can lay a solid foundation for subsequent pedestrian attribute judgment.Five pedestrian key points are defined in our experiments to facilitate the judgment of pedestrian attribute.In order to implement our detection system in real scenarios,we collect videos in street and design a dataset with labels by ourselves,and then we train models and test them in our own dataset.Experiments show that pedestrian detection task and key points task can promote each other.