Research On Object Detection Of Multi-scale Feature Fusion

Real-time detection of vehicles and pedestrians on the road is an important direction of computer vision,the single-stage object detection algorithm has a high real-time performance,but the accuracy is not high.Therefore,this paper proposes a multi-scale feature fusion object detection method based on the YOLOv2 algorithm in the series of YOLO(You Look Only Once)algorithms with the highest real-time performance.The purpose is to make YOLOv2 algorithm ensure the detection speed and further improve the accuracy of object detection.Firstly,in order to make the data in training process more pertinenced,the model selects the most suitable number and size of candidate boxes by clustering the KITTI data sets before the training.Then,in the training part of the network structure,the residual block structure is used to add semantic information,in order to further retain the underlying information and extract the feature description more in line with the object.Finally,in the detection part of the network structure,the feature pyramid network is introduced to fuse the feature graphs of different sizes,so that the high-level feature graphs also have rich semantic information.At the same time,the training was carried out in the way of pre-training,and the final RF-YOLOv2 model was obtained by fine-tuning the trained darknet19?448.conv.23 model.In order to verify the effectiveness of the improved RF-YOLOv2 model,this paper conducted verification on the KITTI data set,and evaluated the performance of the model by comparing various evaluation indexes with the YOLOv2 and YOLOv3 algorithms,and the influence of image scale on the performance of the model was further explored.The experimental results show that RF-YOLOv2 model can improve the accuracy of object detection under the condition of real-time performance.The larger the size of the input image,the higher the detection accuracy will be and the slower the detection speed will be.However,with the increase of the size,the improvement of the accuracy will be weakened and the influence of the speed will always exist.Therefore,when selecting the size of the input image,the actual requirements and the characteristics of the detected object should be taken into consideration.There are 37 figures,4 tables and 52 references in this paper.