As a result of the research, it can be found that it is possible to build a personal navigation system that can localize a person indoors using independent sensors. Similarly, eyt is the y-coordinate measurement error and ephit is the direction measurement error.
Background and Motivation of the Dissertation
In a sense, the give-and-take approach to human-robot collaboration is even more natural than human-to-human collaboration. In the above examples (Figure 1.2), a person has a general plan (today we are going to do some gardening here or we need to level this part of the yard) and shares it with.
Case Study: PeLoTe Project
The common model can only be maintained if all information is reported relative to the given frame of reference. For example, the human device may report to the operator that the corridor in front of me has collapsed, and the operator marks the corridor as a restricted area on the map.
The human entity working at the operational level was also part of the system. PeNa is a wearable sensor system that infers the position of the human device in the shared model.
Main Contribution of the Dissertation
The design of the system is motivated by the PeLoTe project and therefore also by the simulated search and rescue scenario. Using a human's location provides valuable information to the operator coordinating the team.
Author's contribution within research groups
Declaration of previous work
A human team member is referred to as a human entity (HE) and a robot team member is called a robot entity (RE). The setup offers two points of view for the thesis: 1) remote coordination of mixed human-robot teams and 2) cooperation in mixed human-robot teams. The collaboration between mixed human-robot teams is a much broader area of research and is therefore mostly left for future work.
Human team work
In fact, Jones and Hinds noted that the role of the tactical commander in maintaining the common ground during the mission was significant. The commander had the most complete picture of the situation, which he updated if necessary for the team members.
Situational awareness in the traditional sense is completed in team coordination by a supervisor. For example, in an information gathering task (e.g. ) the map is shared among all team members.
Terminology and methodology
- Frame of reference, maps, and localisation
- Continuous localisation
The purpose of the particle filter is to represent the posterior with the particle set. The red and green coordinate axes in the image represent the weighted average of the pose.
Personal Navigation Systems
- Personal Dead Reckoning systems
- Bounded error Personal Navigation Systems for Indoors
- Experimenting with commercial systems
Gravity causes errors in the signal if the orientation of the sensor is not completely known. The transformation of the accelerations to the world coordinate system requires knowledge of the orientation of the sensor. The sources of error in this case are the drift in angular velocities and the error in the initial position of the sensor.
This pattern is caused by the vertical movement of the center of mass, as illustrated in Figure 3.15. With this method, the paper reports an error of up to 5% of the distance walked. In their work, the orientation of the sensor was also used to estimate the movement of the foot in a plane.
Foxlin presents a Kalman Filter-based solution, which estimates the position, attitude and deviations of the sensor. The accuracy of the module is promised to be 2% to 5% of the distance traveled.
- Review of range sensors
- Laser scan matching
- Map-based localisation
An environmental perception (sometimes called extraceptive) sensor measures the current state of the environment. The transformation converts the current scan points to new coordinates defined by equation 3.35. NDT calculates an approximation of the normal distribution in the discretized cells based on the scan points that are in the cell.
The basic assumption is that given two mest−1 dest scans taken of the environment, the tangential direction of the obstacles remains the same in both scans. The common direction is illustrated in Figure 3.25 and can be found by searching for the maximum of the histogram . EKF provides a minimal state variance estimate based on a motion model and a measurement model.
Both the measurement and the map also contain noise, so the result of matching the measurement to the map is also treated as a distribution (that is, the measurement can be taken from an area defined by a mean and variance) . The idea of a probability grid is to represent the posterior of the pose belief over all possible poses in a discretized manner.
In the scheme, the heading can be estimated using a heading filter, which returns an absolute heading, or as a differential change in heading between scans. Basically, the methods presented provide a means to build a personal navigation system that: 1) uses personal dead reckoning and a map, without any environmental sensing sensors; In practice, the third option was chosen for use in the final version, because it turned out to be the most robust at that time.
In addition, later in this chapter a map matching method based on odometry and map matching will be presented, without using lasers at all. The PeNa device includes batteries, power converters, a step length measurement unit (called SiLMU or NUPPU, which are presented in section 4.3.2), an optical gyroscope, a 3DM-G IMU, a compass, a scanner SICK LMS200 laser, a camera and two laptops. A laptop is installed in front and does all the necessary calculations and serves as a screen for the user interface.
The study used two different approaches to estimating stride length: one based on ultrasound and the other based on accelerometers. The accelerometer-based system was chosen for the final demonstrations because there were other systems (namely robots) that used ultrasound devices and these might have caused interference. For heading estimation, PeNa supplies a Hitachi Fiber Optic Gyroscope HOFG-X, a compass and a MEMS IMU from Microstrain (3DM-G).
The HOFG-X sensor was used in the evaluation of chapter 4.3.1, which was used when deriving the laser dead reckoning results (section 5.2.2) for the algorithm presented in section 22.214.171.124. However, the sensor provides a means to study whether a 2D range sensor can be used for personal navigation 1.
Personal Dead Reckoning
- Heading Estimation
- Step Length Estimation
- Upper body location estimation
- Laser-Based Dead Reckoning
- Increasing Error tolerance
The lower image shows the dependence of the measurement on the location (indicated by numbers on both images). The last maximum and minimum are saved and used in the step length calculation. The upper body header is assumed to be fixed when a person has both feet on the floor.
For example, consider the case where the same point is measured in both scans and the distance from the laser origin to the points is 5 m. The method presented in the previous section is good for testing the feasibility of scan matching. In the early phase of the work, the histogram matching from  (based on ) was tested.
The angle between two laser scans is corrected by using the angle histogram of the scan . For example, in Figure 4.11, the maximum of the correlations is slightly between the different step sizes.
- Monte Carlo Localisation Algorithms
What is more important is that the dead reckoning estimation error has a mean of zero (ie the measurement is not biased) and that the variance is known. The general map in the figure is derived from the CAD drawing of the building. Because of this, a measurement should not have much significance associated with particle weight.
The update phase compares the expected measurement with the actual measurement and calculates the probability that the measurement will be taken from the location represented by the particle. Because of this, one measurement (or even a few) should not affect the position distribution weights too much. An example of a distance transformation and a pre-calculated weighting function is presented in Figure 4.22.
It is called topological because the pose update process matches the trajectory of the particles on the map. The particles move in dead reckoning and the update phase only updates the probability, not the position of the particles.
Step length tests
Personal Navigation tests
- Error Estimation
- Dead Reckoning Results
- Map-matching methods
The filled area is the corridor network and the black dot marks the start and end points of the test runs (in most cases). The results of the map-based methods, as well as the laser dead reckoning algorithm (Algorithm 5), are post-processed. The dead reckoning output is obtained by using the translational output of the dead reckoning laser and heading from the heading filter.
As the results show (see Figures A.5, A.6 and A.7), the method performs well in all test sets. Therefore, this method was not used when the lengths of deadlift trajectories were compared. The map used in the previous sections is based on a CAD model of the building.
In all cases there is a significant directional error (e.g. the first data set has a directional error of almost 90 degrees in the first 30 m of walking). For the first data set (upper left corner of Figures 5.8 and 5.9), the algorithm had to be adjusted quite a few times before it was successful (noise parameters and map selection).
Human-Robot Team Tests
- The experimental setup
The autonomy level of the PeLoTe system can vary from direct remote control to full autonomy. An example of the operator GUI and the human entity GUI is given in Figure 5.13. The first experiment (called the semi-nal experiment) was carried out in October in the Physics Building of the University of Würzburg.
Estimating the impact of PeNa on the whole system from the system evaluation is difficult. During the tests, it was found that the 30-minute training session was insufficient for most teams. However, the supervisor was able to track the position of the human from the GUI and could guide the human entity through audio communication.
Overall, all the indicators showed that the PeLoTe system improved the understanding of the situation for the whole team. It is impossible for the operator and human entity to keep track of the events.
Case 1: Correlation in the pose space
Case 2: Combined angle and position correla- tion scan matching
Tables for all MCL runs
Paths of all MCL runs