Research On The Key Technologies Of Autonomous Navigation For Mobile Robots In Unstructured Environments

With the development of science and technology,traditional surveying and mapping has been extended to mobile mapping and real-time kinematic mapping.Especially,robotics represented by unmanned vehicles,unmanned aerial vehicles,unmanned ships and submersibles is a hotspot in the field of surveying and mapping,as well as in natural science and engineering technology.The application of robots involves many fields such as life,medicine,agriculture,construction,military and space exploration.Compared with traditional industrial robots,highly intelligent mobile robots with independent perception and behavior decision have more extensive development potential.Working in unstructured environments,mobile robots have to obtain the ability to perceive the environment in real time and plan an optimal path according to the surrounding environment to complete the task.The key to make robots autonomous mobile is the autonomous navigation technology.The abilities of accurately autonomous localization,environmental perception and path planning are the premise and foundation for the autonomous mobile of unmanned systems such as unmanned aerial vehicles and unmanned vehicles.In this paper,some key technologies of mobile robots autonomous navigation are further discussed and studied,which include real-time visual odometry methods,robust graph optimization based vision navigation methods in dynamic environments,monocular visual inertial navigation methods and global path planning methods.The main works and innovations are as the following:(1)The camera model and relative pose estimation technique are studied.A method for camera distortion calibration based on fundamental matrix is proposed.Based on epipolar geometry and single parameter calibration model,the equation of correspondent points is formed.Utilize a two-step iterative optimization strategy to avoid the problem of instability result from too many unknown parameters.Experimental results show that the proposed method can obtain distortion parameter and the principle point coordinate only by using two images,which is robust to the noise.As for the 3D-2D perspective projection problem,the feature-based method and direct method are compared and analyzed comprehensively in terms of accuracy,computational complexity and robustness.Experimental results show that the accuracy of both methods is in centimeter,while the feature-based method is more suitable for real-time pose estimation in autonomous navigation of mobile robots in terms of robustness.(2)Aiming at the special application scenario of vehicle city navigation,a method of monocular visual odometry method based on prior information is proposed.Based on the assumption of knowing the height of the camera,a robust single scale factor estimation algorithm is proposed to solve the problem of absolute scale ambiguity in monocular vision odometry.Experimental results show that the absolute positioning accuracy of this method is in meters,and the single frame average processing time is about 0.07 s with single thread programming on a laptop,can meet the requirements of vehicle autonomous navigation city in real time.(3)For the problem of real-time precise localization for mobile robots' autonomous navigation,a stereo visual odometry based on the Kalman fusion of optical flow tracking and trifocal tensor constraint is proposed.To speed up the algorithm,image sequences are classified into key frames and non-key frames.Utilize conventional feature point detection and matching method to process the key frames and utilize Lucas Kanade optical flow method to track corresponding feature points in non-key frames.The observation mathematical model based on trifocal tensor constraint between image triples is derived,which is combined with dynamic equation to form the Kalman Filter model.An iterated Sigma Point Kalman Filter is employed to cope with the nonlinear system.To improve the robustness of visual odometry,a RANSAC-based method is applied to pure matching points in motion estimation.Experimental results demonstrate that the absolute positioning error of the proposed method is in meters,and the average processing time of one single frame image is about 0.05 s.(4)In dynamic environments,the moving landmarks can make the accuracy of traditional vision-based navigation worse or even failure.To solve this problem,a robust graph-based vision navigation algorithm with dynamic landmarks is proposed.The motion index is added to the traditional graph-based vision navigation model to describe landmarks' moving probability,changing the classic Gaussian model to Gaussian mixture model,which can reduce the influence of moving landmarks for optimization results.To improve the algorithm's robustness to noise,the covariance inflation model is employed in residual equations.The expectation maximization method for solving the Gaussian mixture problem is derived in detail,transforming the problem into classic iterative least square problem.Experimental results demonstrate that in dynamic environments,the proposed algorithm outperforms the traditional method both in absolute accuracy and relative accuracy,while maintains high accuracy in static environments.The proposed algorithm can effectively reduce the influence of the moving landmarks in dynamic environments,which is more suitable for the autonomous navigation of mobile robots.(5)For the problem of highly mobile robots' real-time precise localization,a monocular visual inertial navigation algorithm based on nonlinear optimization is proposed.To handle multi-rate measurements in visual inertial navigation system,the IMU pre-integration technique is utilized to process inertial measurements.A fast accurate linear initialization method is introduced to estimate the initial scale,gravity direction,velocity and gyroscope and acceleration biases.Two kinds of nonlinear optimization based visual inertial tightly coupled model are built according to whether the map points are updated.The pipeline of sliding window optimization with global map point constraint is introduced in detail.Experimental results indicate that the initialization method can obtain accurate initial states in few seconds.Compared with visual navigation only,the proposed algorithm can obtain the absolute scale and higher working frequency.Compared with classical sliding window methods,the proposed algorithm can effectively decrease the influence of accumulated errors,which validates the correctness and feasibility of the proposed method.(6)To meet the requirements of global optimal and real-time obstacle avoidance in mobile robots path planning,a novel method based on the fusion of improved A* algorithm and dynamic window approach is proposed.Firstly,based on the combination of Manhattan distance and Euclidean distance,a more appropriate heuristic function is designed for A* algorithm.Then a key nodes culling scheme is introduced into the traditional A* algorithm to remove the redundant nodes.Furthermore,an evaluation function considering globally optimal path is constructed.The dynamic window approach based on the evaluation function is applied to perform real-time dynamic path planning,which can guarantee the smoothness of path and the local obstacle avoidance ability,while holding the global optimality of path.Experimental results demonstrate that the proposed method can generate a smoother path,possesses the ability of dynamic obstacle avoidance and guarantee the global optimality of path planning.