Author: Himanshu Bobade
Artificial Intelligence is a data driven technology that is advancing at a rapid pace for the past decade. Certainly, GNSS acts as an important role for global positioning and navigation. Even after achieving the centimeter scale accuracy, GNSS faces challenges of atmospherical noises and inaccuracy due to dynamic objects. Machine learning can help correct the results improving the efficiency. The ways the AI can contribute to the GNSS is set out completely.
Figure-1
Introduction:
What is GNSS?
GNSS stands for Global Navigation Satellite System. GNSS is a constellation of satellites providing signals from space that are transmitted, providing positioning of any object and also the timing data to GNSS receivers. The data received by the receivers is then processed and used to determine the location. GNSS can also refer to augmentation systems. A few key advantages are to have access to multiple satellites with better accuracy, redundancy, and availability at all times.
How does GNSS work?
GNSS works on the basic principles of Physics and Mathematics. Satellite navigation systems function on the laws of physics like Kepler's laws, the Doppler effect, etc., the spread-spectrum technique is used to model the satellite signal allowing to receive poor signals transmitted from the medium-Earth orbits. Scientists and engineers have been working on GNSS systems and as a result, GNSS can now achieve centimeter-level positioning. If the GNSS works so well, then why do we need its Machine learning application.
Challenges in GNSS:
Trilateration is the process of determining the receiver's position on the earth's surface. It involves measuring the distance from the receiver to 3 satellites. In trilateration, the assumption is that the satellite signals always transmit in direct LOS(line-of-sight). But naturally, there are different layers in the atmosphere which may diffract the signals and cause transmission delay. Buildings and obstacles cause NLOS (non-line-of-sight) receptions. NLOS receptions are much harder to deal with due to higher nonlinearity and complexity. And to avoid this, Machine learning plays an important role.
How can we use Artificial Intelligence in GNSS?
Machine Learning uses a dataset to train and create a model and execute it. It can help us mitigate multipath effects and NLOS challenges.
Labeling:
The labeling model will illustrate 3D building, bridges, towers, etc. but might not be able to give us information about dynamic objects like a moving car, trees, etc.
Classes and Features:
Multipath, including carrier-to-noise ratio, pseudo-range residual, DOP, etc. affect variables, and features are selected based on these variables. If one can assess a step deeper into the correlator, the shape of correlators in code and carrier are also excellent features. The comparison between the different levels (RINEX, correlator, and NMEA) of features for the GNSS classifier and concludes that the rawer the feature is, the better classification accuracy can be obtained. The methods of exploratory data analysis can better select the features that are more representative of the class.
Making the best Model:
Machine learning often faces the challenges like overfitting. We don't want our model to overfit our model over the dataset. It should not consider extreme cases, it should be able to generalize the data. Multipath and NLOS might vary in every region. The classifiers trained in India might not work well in the US. This is because architecture of the regions is different.
To determine the positioning, various sensors are used. The inertial sensors perform well in most regions but the MEMS-INS is affected by noise. Consider an example, Sensors like GNSS+IMU are used for walking on a street to the subway station, for walking in a subway station Wi-Fi/BLE+IMU is used, for traveling on a subway IMU is used and for walking in an urban area to the office VPS+ GNSS+ Wi-Fi/BLE+IMU is used for better results. The state-of-the-art detection algorithm can achieve as much as 95% for pedestrians under indoor, intermediate, and outdoor scenarios. This certainly means we can use ML to select right navigation systems intelligently.
When we are supposed to deal with object detection, deep learning is the currently the mainstream method because it generally outperforms ML when two conditions are fulfilled that are data and computation. While ML trains models to fit assumed (known) mathematical models, the trained model of DL is completely data-driven. DL achieves excellent performance due to its superiority in data fitting, if extensive and conclusive data are available.
A DL-trained NN should consider both long-term and short-term challenges, if it can be perfectly designed for the integrated navigation system. A precise positioning system can be attained with the help of Machine Learning. The AI will smartly facilitate the selection of raw measurements and appropriate sensors. Integration Algorithm might be affected by the transient selection of the sensors which are well known as plug-and-play.
VPS:
Visual Positioning system(VPS),developed by Google, can replace the visual corner point detection by the semantic information that detected by ML. For instance, to know where we are and where we are heading, we usually compare landmarks, picturize it and memorize it. In similar way, Machine learning can segment and classify images into various object entities, eg. Building, Laptop, Person, Book ,etc. This object entities can then be compared with the ones on cloud server and provide associated position tag.
Shadow Matching based on LOS and NLOS:
The shadow matching (SDM) is integrated with the SVM classification and a robust estimator. The classification is achieved by following the steps:
Initial approximate solution: The conventional position estimations of shadow matching are weighted least square solution, NLOS probabilities based weighted least square solution.
Particle sampling: The Gaussian distributed particles can be scattered using particle sampling as shadow matching position candidates. A search area is centered at the initial approximate solution. The advantages of this sampling include dropping the number of position candidates and also reducing the running time of algorithm. The demonstration of particle sampling is shown in Fig. 1.
Figure-2
Demonstration of particle sampling
(The source link of the output is given in the references)
Observed satellite visibility: As mentioned earlier, the dynamic objects may cause noises or blocking of the signals. The elevation angles of satellites and building boundaries at the same azimuth angle help predicting the satellite visibility of each particle.
Predicted satellite visibility: SNR measurement and elevation angle of satellite contribute to the conventional LOS/NLOS classifiers. The attempts still face the issue of SNR unreliability and of its changing behavior. Due to these reasons, a multi-featured classification approach is necessary to supply SVM classification.
Conclusion:
Thus, we have learnt a lot about GNSS and the roles AI has in GNSS. The AI will influence navigation research and development. It is highly probable that ML’s development and achievement in the field of navigation will contribute to GNSS technology.
References:
https://satellite-navigation.springeropen.com/articles/10.1186/s43020-020-00016-w
コメント