| With the development of artificial intelligence technology,mobile robots have received widespread attention and applications.Utilizing self-assembled sensors to achieve autonomous positioning accurately and reliably has become a key prerequisite for mobile robots to successfully complete various intelligent tasks.However,the operating environment of mobile robots has characteristics such as large range,complexity,and dynamic changes.Currently,traditional positioning and navigation methods based on manually designed models,such as Inertial Navigation System(INS)and Global Positioning System(GPS),are difficult to apply to all environments and have certain limitations in practical applications.In recent years,data-driven positioning and navigation methods,which do not require manually specified model structures and avoid the problems of traditional positioning methods,have been rapidly developing into a new field.Therefore,this paper takes mobile robots as the research object and aims to improve the positioning accuracy and robustness of the navigation system.Based on deep learning methods,it studies key technologies such as MEMS-based Inertial Measurement Unit(IMU)signal denoising algorithms,lightweight and efficient IMU attitude estimation algorithms,INS/GPS integrated navigation systems,and Light Detection and Ranging(Li DAR)odometer positioning algorithms.The main research contents and innovations of this paper are as follows:1.In response to the actual movement of mobile robots,consumer-grade MEMSIMUs are susceptible to complex and time-varying errors and noise interference,which causes a time-drift problem in positioning results.The error characteristics and signal denoising method of MEMS-IMU devices are studied.By analyzing the error characteristics of MEMS-IMU devices,a hybrid denoising model based on an attention mechanism and convolutional long short-term memory neural network is proposed.First,the MEMS-IMU measurement data is regarded as a type of time-series data.A onedimensional convolutional neural network is used to extract the local key features of the MEMS-IMU measurement sequence.Then,a long short-term memory neural network is used to learn the correlation of different local features on different time scales.To improve computational efficiency,an attention mechanism module is designed to automatically allocate weight coefficients to distinguish the importance of feature data at different time points,achieving precise characterization of the output data of MEMSIMU and thus achieving the goal of denoising.The experimental results indicate that the proposed denoising method reduced the average bias instability and angle random walk error of the original gyroscope data by 57%and 66%,respectively.2.Regarding the current problems of the MEMS-IMU attitude estimation algorithm based on deep learning,such as insufficient consideration of the inherent connection between the accelerometer and gyroscope,large parameter quantity,and poor real-time performance,a lightweight attitude estimation method based on the IMU error compensation principle is proposed.Based on the derived MEMS-IMU measurement model and attitude update model,a lightweight error estimation network model is constructed using depthwise separable convolution,which dynamically regresses and predicts the output compensation component of the gyroscope.Then,based on the principle of inertial navigation,the corresponding attitude angles are calculated by using the error-compensated MEMS-IMU measurement data.The parameters of the error compensation model are optimized and updated to improve the accuracy of attitude estimation by designing the loss function between the estimated and true attitude angles.Experimental results show that the proposed method can not only achieve the precision of the advanced visual-inertial localization system on some test sequences but also has fewer model parameters.3.For the problem of navigation error accumulation when a mobile robot enters a GPS signal interruption environment and the INS/GPS integrated navigation system degrades into pure inertial navigation mode,a deep learning-assisted INS/GPS integrated navigation method is proposed.By deriving the functional relationship between GPS position increments and INS measurements,a GI-NN deep learning model is designed to predict the interrupted GPS position increment data.GI-NN combines onedimensional convolutional neural networks and gated recurrent unit neural networks to extract spatial features and temporal information from the IMU signals.A relationship model between attitude,specific force,angular rate,and GPS position increment is established,and the mobile robot’s motion state is dynamically estimated using current and historical IMU data.Experimental results show that compared to traditional machine learning algorithms,the proposed method can provide more accurate and reliable navigation solutions in GPS interruption environments.4.A deep learning-based end-to-end LiDAR odometry method is proposed to address issues such as the traditional LiDAR odometry’s reliance on manually designed point,line,and plane features and the complexity of the pose estimation process.A deep recurrent convolutional neural network is used as the basic framework for pose estimation,with a LiDAR feature encoder and odometry network module constructed separately.In the LiDAR encoder,a multi-channel,multi-scale convolutional neural network is designed to automatically extract key features from LiDAR point clouds.Meanwhile,a channel attention mechanism is embedded to enable inter-channel information exchange.The odometry module is constructed using gated recurrent unit neural networks and fully connected neural networks to further learn the motion relationships across time scales.By training and learning the neural network’s weights and biases,the corresponding translation and rotation components are regressed and predicted,thereby avoiding the complex matching and optimization steps of traditional methods and improving the accuracy and robustness of LiDAR odometry. |
PDF Full Text Request