Research On Small-scale Pedestrain Detection Algorithm Based On Yolov3-tiny

By using computer vision technology,pedestrian detection technology can detect whether there are pedestrians in an image or video sequence and give pedestrian accurate locations.Combined with pedestrian tracking,pedestrian recognition and other technologies,pedestrian detection technology can be applied to intelligent video monitoring,human behavior analysis,intelligent transportation,vehicle assisted driving system,intelligent robots and other fields.In some application scenarios,small-scale pedestrians account for a large proportion,it is easy to cause information loss,and existing pedestrian detection algorithms have low detection accuracy for small-scale pedestrians.The main work of this thesis is as follows:(1)In view of the shortage of small-scale pedestrian proportion in the existing pedestrian datasets,this thesis proposes a pedestrian dataset SPD2019,the size of pictures is 1600*900 pixels.The dataset is a supplementary training set for the existing datasets,including 1200 training pictures and 300 test pictures.The existing datasets mainly relates to life scenes or street scenes taken by car cameras.The shooting angle of SPD2019 is located at high places with depression angles,and the shooting scenes are scenic spots and streets,the proportion of small-scale pedestrians(the height of pedestrians is less than 30 pixels)accounted for 55%.Using YOLOv3-Tiny as the test algorithm and SDP2019 dataset as its supplementary training set,After the simulation experiments,the average accuracy of small-scale pedestrians on the existing dataset(COCOPerson,City Person,Caltech,WiderPerson)and SPD2019 test set improved by 3.4%and 64%,respectively.(2)YOLOv3-Tiny algorithm reduces the size of the image before the image is input to the network,resulting in serious loss of small-scale pedestrian information.To solve this problem,this thesis presents a CEYOLO(crop and enlarge Yolo)algorithm which can crop the original picture.The CEYOLO algorithm adds a module which can predict and crop the small-scale pedestrian area in the original image,and re-enter the cropped area into the network for detection.The experimental results show that compared with YOLOv3-Tiny,CEYOLO’s average accuracy of small-scale pedestrian has increased by 35%,and the overall average accuracy of pedestrians has increased by 14%.(3)Finally,based on the improvement of dataset and algorithm,the trained CEYOLO model is successfully implemented and tested on the embedded platform Jetson Nano,and the video decoding and model reasoning are accelerated.On the embedded platform,the average detection time of each picture is 0.29s,and the average accuracy is 0.702,which achieves the expected detection target.To sum up,this thesis labeled SPD2019 dataset and proposed CEYOLO algorithm so as to improve the average accuracy of small-scale pedestrian detection.