Research On Real-time Object Detection And Tracking For Mobile Robot Based On Deep Learning Method

In recent years,the breakthrough of deep learning technology in the field of computer vision has made its research more and more attention.At present,the foundation of deep learning-convolutional neural networks still has problems in practical applications such as poor generalization ability and high computational complexity.This paper focuses on the use scenarios of mobile robots and conducts research based on the application of deep learning in target detection and tracking.The main work is as follows:Aiming at the problem of multi-scale target detection,in order to improve the accuracy of multi-scale target detection,this paper uses a multi-scale feature of convolutional neural network to study a multi-scale fusion method based on feature conversion.First,in view of the limitation of the computing power of the mobile robot platform,this paper reduces the original YOLOv3 network Darknet-53 network to obtain the YOLOv3-tiny network,on this basis,the STDN scale conversion method is used to scale the shallow network and scale the deep network.Training,and testing the detection accuracy on the COCO dataset,and verify the real-time performance through the cross-compilation on the Jetson Nano embedded platform.The results show that the method in this paper guarantees the real-time detection accuracy and originality on the low-power GPU platform.The YOLOv3 network is not much different.Aiming at the problem of single-target real-time long-term tracking,in order to ensure the accuracy and real-time of tracking at the same time,this paper studies an online tracking algorithm by combining the deep learning tracking algorithm with the traditional long-term tracking algorithm.First,this paper compares and analyzes the classic long-term tracking algorithm TLD and the real-time deep learning tracking algorithm GOTURN,and finds that the two have complementary advantages and disadvantages.On this basis,this paper proposes an improved algorithm GOTURN-ld combining their advantages.The improved YOLOv3 network is introduced as an improved algorithm detector,and the accuracy of algorithm tracking is further improved on the basis of ensuring the real-time performance of the algorithm.In this paper,the improved tracking algorithm is applied to the mobile robot system for indoor scene experiments,and the main framework of the system and the transformation of data and instructions are designed.The tracking performance of the improved algorithm is verified by testing the mobile robot under different lighting conditions.