Research On Pedestrian Detection Of Small Target Based On Convolution Neural Network

Pedestrian detection is an important sub-topic in the field of object detection.In emerging applications,whether for video or image,the pedestrian scale is very small due to the distance problem,so conventional pedestrian detection methods can hardly detect such small target pedestrians,and even the test results based on some deep learning models are unsatisfactory.In addition,the detection of small target pedestrians is very difficult due to factors such as camera movement,similar color blending of the scene,lack of light,or shadow interference.Nowadays,the effect of deep learning technology on pedestrian detection is much better than that of traditional machine learning and other image processing methods.In particular,there are many pedestrian datasets used to compare the detection results.With the deepening of researches,such influencing factors as illumination,deformation and occlusion of similar images have been partially solved,but the small target pedestrian detection is still a difficult point in pedestrian detection.Based on this problem,this paper studies the detection of small target pedestrian based on convolutional neural network method.The main research contents include:1.The models of pedestrian detection based on convolutional neural networks have been studied in recent years.Two models are mainly introduced.One is based on candidate region extraction models such as RCNN,Fast-RCNN and Faster-RCNN,and the other is based on the overall image.For example,Regression models such as YOLO and SSD.By comparing various aspects of these convolutional neural network models,using Faster-RCNN,YOLO and SSD models to detect general-scale pedestrians in the image,and finally analyzing the detection results,the rationality of using YOLO as a base model to improve is demonstrated.2.In this paper,based on the original YOLO model,an improved convolution neural network model YOLO-K is proposed.The improved method is to add a new kronecker upper sampling layer and a sub-region extraction layer before the feature extraction layer of the model.Then the experiment modify the relevant parameters of the detection part of the model.In this paper,we use the image data of UAV to acquire experiment independently,and detect the small target pedestrian in this data set.The final experiment uses the improved convolutional neural network model YOLO-K,the original YOLO model and the SSD model to detect the pedestrian targetin the aerial aerial image of the UAV respectively.The false positive rate and missing detection rate are taken as the indicators of the comparison.The results show that the improved model in this paper can detect the smaller target pedestrians well,especially the recall rate has been greatly improved.