The basic technology
Therefore, the main routes are given in our case by piecewise-arc approximation of the full trajectory, and the altitude features of the terrain should probably be taken into account by extensions of the planar segments.
The driver's navigation program is based on a real-time autonomously deployable positioning system (RTLS) operating according to the ToF (Time of Flight) principle and organized with useful data retrieved locally on the host vehicle - similar to GPS systems. It is assumed that the vehicle is also equipped with inertial sensors capable of estimating the distance traveled in one way or another.
- The ratio of the accuracy of the individual measurements of the inertial navigation and the guided navigation with RTLS is taken into account by the choice of the Kalman coefficient, whereby it is allowed within a considerably wide range so that the inertial accuracy can be a multiple lower, for example in the event of mechanical interference with the movement.
- Each arc of the main master route is a driving iteration during which the vehicle is not conditionally controlled by the driver, but can and should be intercepted when the driver detects obstacles and/or other unplanned disturbances. In such a case, logically, the driver program switches to driving the vehicle by an alternative route (standing still, among other things, is considered to be an en-route), and the master route is held unchanged in the background.
- The lengths of iterations, as well as details of route approximation, are chosen to be reasonably small, limited by the speed and performance of measuring systems, the need for movements with acceleration (transient movements), and the characteristics of the vehicle.
- Route guidance is accomplished by setting a vehicle task in each iteration, described by at least two parameters: angle and arc length. Temporal parameters such as speed, acceleration, etc. can act as additional parameters, but it is the temporal characteristics that are secondary and can be defined independently to provide the track geometry.
- The duration of the iteration for the driver is arbitrary, the iteration ends with notification from the vehicle that the task has been completed. It is worth noting here that if the vehicle performed the task perfectly, without any executive error of both external and intrinsic origin, then effective driving is achieved; but practically it is impossible, and hence the control and correction of such error is the content of driving, which in the proposed solution is done by the orienting navigation via RTLS system.
- At the end of the iteration, the driver takes a measurement in this system and receives, as a result, a vector of correlated distances from the radio tag installed on the vehicle to the radio beacons, previously known location in some Cartesian coordinate system. For this result to be useful, a multilateration of the vector of distances must be performed and an estimate of the coordinates of the tag and the vehicle itself must be obtained. The exact solution of the multilateration problem with a stochastic distance input vector and an arbitrary number of elements of such a vector, i.e. the number of visible beacons, is practically impossible to implement.
At the moment of vehicle respawning at the driving site, full-fledged position determination is naturally possible when four or more radio beacons are visible, but in continuation of driving the method is already stable to changes in their number:
- normally operable with more than two distance-vector elements
- with two - if there are no significant drops in height
- with one - for some very limited time
Note that since the RTLS measurement only contains information about the translational degrees of freedom, but not about the rotational ones, it requires the assumption of an arc-shaped path to calculate the correction. What is notable about the driver program, however, is that the continuously calculated corrections for the basic transformation matrix and the corrective action, which are passed to the TC problem at the next iteration, are assumed to be logically independent. Moreover, the corrective action is composed of two additive components: a differential component that compensates for the error estimated from the previous driving iteration, and an integral component that compensates for the total deviation of the vehicle from the master path. This corrective action is itself additive to the underlying routing in the iteration. Therefore, at the beginning of each following iteration, the program driver calculates the integral route correction, adds it to the differential correction calculated at the end of the previous iteration, then calculates and adds to them the main route, thus forming the task of the vehicle, gives this task to the vehicle, waits for its execution and the actual inertial data about it, during which it analyzes and monitors the environment, receives the notification of task completion, then performs measurement in RTLS and multilaterates it then calculates the main transition of the transformation matrix and calculates another differential correction, comparing the position expected from the task and inertial data with the results of the multilateration and corrects the transformation matrix.
This is where the iteration is finally completed and the cycle is repeated again. It is explicitly allowed that the vehicle may intentionally change the task it has accepted for one reason or another by giving a report of the results of the movement that has taken place, it must be said that the task data used by the driver for further calculations is and only is taken as the data from this report. To analyze and control the environment during each route iteration, which firstly includes obstacle tracking, detection, and measurement, perception and recognition technologies using neural networks or similar structures that receive only a video stream from one or more cameras installed on the vehicle as input data will be used.
The idea and technology of intelligent neural networks based on the division of feature space into corresponding classes and subclasses by means of closed solver boundaries rather than hyperplanes is being developed, investigated and tested, even taking into account the so-called kernel trick.
The key content in the realization of such an idea is the necessity to train such a network with the help of the complex branching evolutionary genetic algorithm, which forms not one successful vector of genes (genotype) but many of them as a result.
The advantages of such a classification and recognition technology come down to the potential overcoming of two well-known problems of existing similar solutions simultaneously:
- the Open Set Recognition problem
- the Catastrophic Interference problem