Real-time Object Detection Based On Improved YOLO Algorithm

In recent years,the rapid development of object detection is a hot research topic in computer vision and pattern recognition.Especially in the context of the big data era,many object detection algorithms based on deep learning have emerged.After weighing the detection speed and detection accuracy,this thesis chooses to study the real-time object detection based on the one-stage YOLO algorithm.The main research content and innovation have the following four aspects:1.Improved the network structure of the YOLOv3 algorithm.The proposed video-YOLOv3 model combines 4 times down-sampling and 8 times down-sampling feature maps into the third detection layer,and its m AP value is 2% higher than the original YOLOv3 algorithm.2.Improved the network structure of the YOLOv4 algorithm.On the basis of the YOLOv4 algorithm,the feature fusion method of point-by-bit addition is adopted,and the YOLOv4-416 model and the YOLOv4-85 model are proposed.The AP value of each category has been improved,and the m AP value of the YOLOv4-416 model reached 80.32%,the m AP value of the YOLOv4-85 model reached 81.97%.3.This thesis optimizes the real-time detection mechanism and improves the effect of real-time detection.Taking full advantage of the rich semantic information between adjacent images in real-time detection,a pixel threshold method assisted by frame skipping is proposed.The pixel threshold method is to calculate the pixels of the continuous image in the RGB color space,and then determine whether to re-detect with the detector,and add frame skip detection to improve the real-time detection effect.The pixel threshold method of adding frame skipping makes the frame rate become about 2?4 times of the original.4.In the research of real-time object detection,in order to further improve the pixel threshold method,an adaptive prediction algorithm is proposed.Compare the related images in the HSV color space,and calculate and process the pixels of the related images through the established mathematical model.The adaptive prediction algorithm improves the sensitivity to pixel changes,and the frame rate becomes about 2 times the original.