Multi-sensors Fusion For Simultaneous Localization And Mapping Of Outdoor Delivery Robots

Outdoor delivery robot is a new type of service robot in recent years.It aims to solve the problem of “last mile”in the last step of logistics allocation.In recent years,major internet companies such as JDcom and Alibaba have been developing outdoor delivery robots whose main technology is inherited from the driver-less car technology but there are also some differences.On the one hand,because of the limited space,delivery robots cannot be equipped with high-power and high-performance CPU and GPU.On the other hand,in order to enable the delivery robot to be commercialized and control its use cost,high-precision sensors like 64-line 3D laser radar cannot be used.How to realize the function of independent distribution of outdoor delivery robots under the condition of limited computing resources,limited precision sensors and limited vehicle cost is requirement for higher technology.High-precision and high-robust positioning capability is the basis of mobile robot autonomous navigation.Simultaneous Localization and Mapping(SLAM)is often a means for mobile robots to achieve high-precision positioning.For example,indoor service robots often use 2D lasers for SLAM,whose effect is remarkable indoor,however,under outdoor conditions,due to road surface bump inequality,2D lasers often measure false laser spots,resulting in positioning failure.Outdoor unmanned vehicles often use GPS and 3D laser to achieve high-precision SLAM,but this technology is difficult to commercialize due to the high cost of 3D lasers.Therefore,this paper hopes to solve the problem of outdoor positioning of delivery robots based on visual simultaneous localization and mapping(V-SLAM).Visual sensor with low price and rich characteristic information has inherent advantages.However,single visual sensor has unstable outdoor performance,such as when the light changes dramatically or a large dynamic object appears,leading to the failure of positioning.The problem of scale uncertainty is also often criticized when SLAM using monocular vision.In addition,in the outdoor environment on a large scale,in order to eliminate the accumulated error,the loop-closure detection and posture correction ask for more computing resources.They are all unfavorable to the application of pure VSLAM in outdoor delivery robots.In view of the above problems,this paper uses GPS,IMU and vision sensors to achieve tight integration of simultaneous mapping and positioning.GPS,IMU and vision sensors have their own advantages and disadvantages: GPS measurements do not have cumulative errors,but the condition of large positioning errors often occurs due to interference from high-rise buildings or shades;IMU have higher accurate in the short term but long-term deviations will be larger;The information of visual sensor is rich but its dynamic characteristics are weak.Therefore,the main work of this paper is to realize SLAM fusion three sensors through the following process.Firstly,for the problem that GPS is vulnerable to environmental interference,this paper defines the GPS confidence function,which can represent the strength of GPS signal constraint;then,the k-means semi-supervised aggregation method is used to identify the parameters of GPS confidence function which can be used to classify new GPS samples.Experiments show that this method is more practical than the method of setting the discriminant threshold manually.Then,for the scale uncertainty of monocular image,the uncertainty of IMU bias and the transformation uncertainty between the geographic coordinate and the map coordinate,this paper conducts an in-depth research on the initialization calibration for three sensors.Through the theoretical derivation,the state variables are decoupled and estimated to obtain their initial values step by step,which lays a foundation for the fusion SLAM.The algorithm proposed in this paper is based on the monocular ORB-SLAM,adding IMU and GPS constraints.To solve the problem of the real-time degradation when mapping caused by the increase of state variables,the mechanism of dynamic local window size is proposed.It can reduce the time consumption and save computing resources while ensuring the accuracy of SLAM.Finally,the paper verifies the accuracy and time-consuming of our algorithm by experiment.In terms of accuracy,compared with the original ORB-SLAM,the proposed algorithm eliminates the cumulative error and restores the true scale.Compared with the loose fusion positioning method,it has strong anti-interference ability to unstable GPS signals;In terms of time-consuming,our algorithm on both threads of the mapping and pose graph correction has been greatly reduced respectively,reducing the requirements on computing resources.