Research On Pedestrian Detection Based On Deep Learning Network

Pedestrian detection technology,as one of the key research topics in computer vision,has become a research hotspot due to its extensive application value.At the same time,due to the non-rigid characteristics of pedestrians and the complexity of the real scene,pedestrian detection is a difficult point.In recent years,the rise of deep learning has promoted the development of artificial intelligence,which has led to the rapid development of pedestrian detection.Although the detection model based on deep neural network develops rapidly,there are many problems in the practical application.For example,in deep network training,a large amount of labeled data is needed to obtain a detector with good performance,and the process of data annotation usually takes a lot of time.In addition,the detection model based on the region proposal has high accuracy,but the detection speed is slow,which cannot meet the real-time requirements.Although the model based on regression method can provide real-time detection speed,its accuracy is not satisfactory.In this paper,the existing problems in pedestrian detection model based on deep learning network are studied as follows:(1)Since the training of pedestrian detection method of deep neural network needs a lot of marked data,overfitting of network model can be effectively prevented.In this paper,a structure of Generative Adversarial Networks(GAN)is used to enhance pedestrian detection data.Firstly,the random noise is used to cover part of the pedestrian bounding box,and GAN is trained with the image with the noise box and the original image as the corresponding label,so that the generator can learn to synthesize the pedestrian image in the noise box.The discriminator receives a pair of images as input,from which it learns to distinguish the image pair of real image and noise image,or the image pair of generated image and noise image.Visually,this method can generate pedestrian images compatible with the background,and use the enhanced database to train the detector,which improves the accuracy of the detector to a certain extent.(2)Aiming at the problem that the detection model based on deep neural network has a long detection time and cannot meet the practical application.We made full use of the advantages of YOLO,and unified the determination of regional proposal box,feature extraction,target recognition and positioning into the same convolutional neural network,improved tiny-yolov3 and applied it to pedestrian detection.We deepened trunk network of the Tiny-yolov3,in order to better extract the features of pedestrian,increased 3 convolutional layer on the basis of Tiny-yolov3,and introduced the convolution kernel of 1×1 in the network to increase the nonlinear activation function.It can be seen from the experiment that the improved model has a certain improvement in detection accuracy compared with the basic model,moreover,the accuracy and efficiency of detection are well balanced,which can meet the needs of practical application.