Research On High Speed Tracking Algorithm Of Mobile Robot Based On Convolutional Neural Network

In recent years,mobile robots have been widely used in transportation,emergency rescue,urban security,military operations and other fields,showing the development trend of intelligence,autonomy and high reliability.As a research hotspot in the field of mobile robots,visual tracking technology can greatly improve the environmental perception ability of mobile robots and the response speed of mobile robots under the extremely high requirements for the practicability and accuracy of detection in the complex working environment of mobile robots.Robots are able to make decisions quickly and complete the tasks communicated.To solve a series of tracking problems caused by the target being occluded by the environment and losing the target when the mobile robot is tracking the target,the tracking target being lost due to the low detection accuracy when the target moves at high speed,and the tracking fails caused by the difficulty in expressing the characteristics due to the small proportion of the viewing angle when the target distance is too far,a high-speed tracking algorithm based on convolution neural network is proposed by using the embedded platform,which is of great significance for the expansion and application of the target tracking algorithm on the mobile robot platform.The main research work is as follows:(1)By analyzing the related algorithms of convolutional neural networks,combined with the analysis and comparison of the principles of target detection algorithms in relevant literature,as well as the requirements of real-time detection accuracy and speed in the application field of mobile robots,and taking into account the premise of embedded systems,the research uses YOLO target detection algorithm.It is proposed to use Varifocal Loss to replace the original Focal Loss to train a dense object detector to predict IACS,and use Dropout to reduce overfitting by ignoring half of feature detectors in each training batch.The reliability of the YOLO target detection algorithm is verified by visualizing the features of the model during training,observing the training degree of the training model and the feature extraction of the network layer in real time,and analyzing the loss function curve and the comparison of related training parameters after the training.(2)Research and analyze the relevant theories of target tracking,and propose the integration of the YOLOv5 target detection algorithm based on the Deep SORT tracking algorithm.By constructing the target matrix between different frames in the tracking,by calculating the minimum distance between the matrices,the target matching is established,and at the same time,the future frame state of the target is predicted,and finally the fast detection and tracking effect of the target is realized.After testing,the algorithm achieves high tracking accuracy and meets the requirements of high-speed tracking.(3)Expand the embedded platform NVIDIA Jetson TX2 on the mobile robot platform to realize the optimal target detection and target tracking algorithm,and use the Tensor RT model inference acceleration and optimization model to improve the utilization of GPU,in order to realize data processing on the mobile terminal,greatly reduce the time loss caused by data transmission,and eliminate the impact of the environment on data transmission.It is verified by experiments that the real-time tracking performance of the algorithm reaches 22 FPS,which realizes the purpose of high-speed tracking of the target by mobile robot.