Research On Loopback Detection Method Of YOLO-NKLT Vision SLAM

With the rapid development of unmanned driving,intelligent robot,virtual reality and other fields,visual SLAM technology has also made great progress.As an important part of solving the cumulative error of visual SLAM,the existing loop detection methods based on deep learning have shown relatively excellent performance.However,in the case of complex outdoor scenes such as obvious lighting changes or occlusion objects,the existing visual SLAM loopback detection method based on deep learning does not make good use of the semantic information and scene details of the image,and has problems such as robustness and poor real-time performance.In this paper,the YOLOv5 network model with improved loss function is adopted to obtain image features with semantic information,which makes the extracted features more suitable for loopback detection in complex scenes.In order to improve the real-time of closed-loop detection,a KLT dimension reduction method based on non-dominated sorting is proposed.The details are as follows:(1)In view of the insufficient use of scene details and semantic information in current loopback detection methods,this paper improved the loss function of YOLOv5 to make it more suitable for loopback detection characteristics.In this method,the YOLOv5 network with excellent performance in speed and accuracy is used for feature extraction,and the loss function of the YOLOv5 network is improved to increase the measurement of the center distance and the calculation of the overlap ratio parameter(CIOU),so as to effectively improve the accuracy of the detection of occluded objects.A new loop detection method is established.Examine two common sets of loopback detection data: the New College dataset and the Nordland dataset with more complex variations such as lighting.Experiments show that the method proposed in this paper can detect the closed-loop well,and effectively improve the accuracy of the loopback detection in complex outdoor scenes,and has excellent real-time performance.(2)Image complexity is high in large and complex scenes and the feature vector extracted in the loop detection has higher dimension and takes longer time.Based on this A method of NKLT dimension reduction based on non-dominant sorting is designed.In this paper,the improved NKLT method is used for dimensionality reduction processing and tensor quantization of the acquired image information to avoid the destruction of the structural relationship between the spatial structure and data in the vectorization process.The spatial and band information is considered as much as possible to retain the data information and potential information to the greatest extent.At the same time,the amount of data needed to be processed is greatly reduced by dimensionality reduction,so as to improve the real-time requirement of the whole new closed-loop detection method.The experimental results show that the dimension reduction method in this paper can achieve better results.Finally,the YOLO-NKLT closed-loop detection method proposed in this paper is compared with six other closed-loop detection methods.Experimental verification shows that the method proposed in this paper can achieve better comprehensive performance,higher average accuracy and stronger real-time performance in complex scenes.