Research On Vehicle Detection Algorithm Based On Deep Learning

With the explosion of the number of cars in the world,a series of tra ffic problems and environmental problems have come one after another.In order to better cope with these problems,the intelligent transportation system(ITS)came into being.The intelligent transportation system contains many subsystems.As one of the basic subsystems in intelligent transportation system,vehicle detection is of great significance to its development.Traditional vehicle detection methods,such as the frame differential method,background subtraction method,etc.,are often unsatisfactory.As deep learning re-emerged,it was quickly applied to various fields.This paper combines deep learning with vehicle detection to provide a safer and more efficient intelligent transportation system.This paper firstly studies the classic vehicle detection algorithm and object detection algorithm based on deep learning,introduces the principle in detail,and analyzes their advantages and disadvantages,and then completes the following work:1)Bottom-up path augmented RetinaNet for dense object detection,PA-RetinaNet,is proposed.In order to solve the class imbalance problem that hinders the accuracy improvement of the one-stage object detection methods,this paper introduces a Class-Imbalance loss function,which makes the model pay more attention to t he hard example.At the same time,a bottom-up path enhancement module is added,which shortens the path of information flow and makes the low-layer information easier to propagate.PA-RetinaNet for general object detection is verified by MS COCO dataset set by Microsoft.Experiments have shown that PA-RetinaNet is 4.3 % higher than the state-of-the-art two-stage object detection method,while the speed is similar to the existing one-stage object detection methods.2)A vehicle detection algorithm based on transfer learning is proposed.As a basic part of the intelligent transportation system,vehicle detection provides guarantee for subsequent vehicle counting,vehicle speed measurement,traffic accident identification,and traffic flow prediction.In this paper,the transfer learning method is used to transfer PA-RetinaNet,which is a general object detection,to the vehicle detection field,which saves a lot of training time.The UA-DETRAC dataset is used to fine-tune the PA-RetinaNet that has transfered to the vehicle detection field.The experimental results show that compared with the existing methods,our algorithm does not lose the detection speed,and the detection accuracy is better.