Design And Implementation Of Object Detection Algorithm Based On YOLO

Object detection is a fundamental and difficult subject in computer vision for a long time.Its task is to detect whether the input image contains a predefined target category and output the predicted location and category information of the target.With the rapid development of artificial intelligence,object detection algorithms are also rapidly updated and iterated.However,as the application requirements become larger and larger,the target scene becomes more complicated,and the object detection algorithm also has many limitations.This paper proposes an improved YOLO-bal algorithm based on the characteristics of YOLO series,such as insufficient recognition of small targets,high algorithm complexity,and slow running speed.Based on this algorithm,a set of intelligent camera target detection edge computing system is designed,and the application verification of the algorithm is provided,which provides research ideas for the intelligent camera target detection edge computing system.The main contents of this paper are as follows:(1)Firstly,we review the current status of object detection algorithms.Several mainstream object detection algorithms are introduced.Then the basic theoretical knowledge and network structure of convolutional neural network are explained.Finally,the network structure,loss function,and key technologies of the YOLO series are introduced in detail.(2)Aiming at the natural defects of YOLO series algorithm detection and identification in small targets,an improved YOLO-bal algorithm is proposed,which adds a 1 × 1 convolution layer and improves the network structure;balance and enhance the original multi-scale features,Weighted average of deep and shallow features,so that each layer of features can make full use of the information of the feature maps of other layers for target detection;for the case where the original loss function is not sensitive to the position of the prediction frame,Gio U is used as Loss function.Through multiple comparison experiments,the improved model performance has been enhanced.On the COCO data set,when Io U is equal to 0.5,m AP reaches42.2%,and the speed reaches 49.0FPS.At the same time,the model size also becomes more small.(3)A set of object detection edge systems for running YOLO-bal algorithm was implemented.And then taking the video images collected by the ordinary camera as input,JETSON AGX XAVIER was used as the development platform of the YOLO-bal algorithm to build an edge computing system.Finally,the designed system is verified and tested.The experimental results show that the YOLO-bal algorithm we designed can achieve 42.1% m AP in XAVIER,while meeting the real-time requirements,reaching 48.8FPS.