CN102401656A - A Position Cell Navigation Algorithm for Biomimetic Robot - Google Patents

Abstract

The invention discloses a positional cell bionic robot navigation algorithm, A1, firstly, a wide-angle camera of a robot takes a panoramic picture, converts the panoramic picture into a gradient map, and then, a Gaussian difference filter is used for filtering the panoramic picture to detect characteristic points; carrying out logarithmic polar coordinate transformation on local areas near the characteristic points, wherein the characteristic points correspond to specific road signs; a2, acquiring an angle relative to the true north direction, i.e. azimuth, provided by a compass, for each landmark; the azimuth angle and the road marking jointly determine the position cell at the current moment; a3, determining a transfer matrix by the position cell at the current moment and the position cell at the previous moment; when the robot is exploring in an unknown environment, a transfer matrix is continuously formed, and the transfer matrix forms a cognitive map; a4, jointly determining a motion transformation matrix by the transfer matrix and the cognitive map, wherein the motion transformation matrix is responsible for sending out a motion command; the robot can be applied to intelligent ground cleaning robots, battlefield search and rescue robots and the like.

Description

A Position Cell Navigation Algorithm for Biomimetic Robot

technical field

The invention relates to a position cell bionic robot navigation algorithm. Development of an Autonomous Robot Navigation Algorithm Based on Navigation Strategies of Mammalian Hippocampal Place Cells. It can be used for intelligent cleaning ground robots, battlefield search and rescue robots, etc.

Background technique

Mammals (eg rats, humans, etc.) spend a significant amount of time moving from one location to another. Purposeful movement requires the encoding of motivation, space. The encoding of space and motivation in animals is robust Research on its mechanism has important implications for the field of artificial intelligence.

"Cognitive maps" refer to the way animals internally represent spatial information. A cognitive map can be represented by a series of landmarks. In the process of tens of seconds to a few minutes, you can be familiar with the environment and remember it accurately. "Place cells" are pyramidal neurons in the hippocampus that fire at a high rate when the animal is in a certain local position in the environment, and almost zero when leaving this position. Place cells have the following properties:

1. In a new environment, place cells can be quickly established;

2. The position field (the position corresponding to the position cell emission) is stable, and the same position cell encodes the same position, and it is still stable even after several months;

3. The position field is not tightly dependent on visual information and is also effective in dark environments;

4. The displacement and rotation of distant landmarks cause the displacement and rotation of the position field;

5. The same place cells may be emitted in different environments, with completely different place fields;

6. Place cells are modulated by head direction cells.

Contents of the invention

According to the place cell encoding strategy in the brain hippocampus, the present invention provides a place cell bionic robot navigation algorithm. To achieve the above object, the present invention is achieved by taking the following technical solutions:

Feature-based road sign recognition. Firstly, the camera is used to obtain the feature points, and secondly, the feature points are used to obtain the position information. A network of landmark orientation angles is established by adding feature points and location information. Images were acquired with a low-resolution panoramic lens. The image is gradientized to remove brightness interference. Gradient image is convolved with double Gaussian difference operator for feature recognition. Signpost neurons learn signposts. For each signpost, the relative angle to true north is read by the compass. The visual system of this model provides both "what" and "where" information.

The information fusion of road signs and direction angles produces place cells, and new positions that appear during exploration are encoded with new neurons. At a given location, multiple place cells are active and co-localized. The density of landmarks is related to the location of the environment where the robot is located. More landmarks are learned where the position-angle positions change rapidly, such as walls and doors. When the entire environment is learned, the environment is completely covered by place cells, each of which corresponds to a corresponding location. Place cells determine where the robot is located.

Transition matrix encoding. For planned navigation tasks, certain trajectories must be completed. These trajectories can be represented by sequences of location points. The entire matrix (transition matrix) from the current position to the next position can be used to represent this trajectory. Impossible portions of trajectories can be removed from the matrix to reduce resource usage.

When the transfer matrix is established, a cognitive map is formed. Cognitive maps and transfer matrices together form the motor transformation matrix. The motion transformation matrix is responsible for issuing commands. For example, transfer cell AB is generated from location A to location B. This transfer cell is associated with the direction from A to B.

Autonomous robots, like humans, should also be designed with motivation. Motivation can be a task to be completed, looking for a power source to recharge when the battery is low, or returning to a fixed place after completing a task.

The invention discloses a position cell bionic robot navigation algorithm. Development of robot navigation algorithms based on mammalian hippocampal place cell-based navigation strategies. It can be applied to intelligent cleaning ground robots, battlefield search and rescue robots, etc. It can automatically identify unfamiliar environments, and has the advantages of self-organization and self-adaptation without human intervention.

Description of drawings

Fig. 1 hardware block diagram of the present invention;

Fig. 2 algorithm flowchart of the present invention;

Fig. 3 is a schematic diagram of formation of landmark-azimuth cells of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with specific embodiments.

As shown in Figure 1, the hardware based on the location cell bionic robot navigation algorithm of the present invention includes: a wide-angle camera for obtaining external images; a DSP chip for performing relevant algorithms for robot learning and cognitive environment; driving wheels, Drive the robot to move according to the motion command sent by the DSP chip.

Biological place cells are only related to the location of the creature, and have nothing to do with the direction and speed of its movement. In order to mimic biological place cells, place cells are modeled with variables in the software environment of the DSP.

As shown in Figure 2, firstly, the robot’s wide-angle camera takes a panoramic photo, and converts it into a gradient map in order to remove the interference of brightness, and then filters it with a Gaussian difference filter to detect feature points. The logarithmic polar coordinate transformation of the local area near the feature point can improve the recognition accuracy of small rotation or scale changes. Since these feature points correspond to specific landmarks, the variables corresponding to these feature points are called landmark cells.

For each waypoint the angle (ie azimuth) is obtained relative to the direction of true north, which is provided by the compass. A 360-degree field of view is encoded by N _Azm azimuth cells. The azimuth and the road marking jointly determine the position cell at the current moment.

For planned navigation tasks, certain trajectories must be completed. These trajectories can be represented by sequences of location points. The entire matrix (transition matrix) from the current position to the next position can be used to represent this trajectory. Impossible portions of trajectories can be removed from the matrix to reduce resource usage. The place cells at the current moment together with the place cells at the previous moment determine the transfer matrix.

When the robot is exploring the unknown in the environment, it continuously forms a transfer matrix, and the transfer matrix forms a cognitive map. The transfer matrix and the cognitive map jointly determine the motion transformation matrix. After outputting motion commands for a period of time, when the transfer matrix is established, it forms The so-called cognitive map. A cognitive map can be considered as a node composed of transfer vectors and boundaries. The node's own transfer weight is set to 1, and the transfer weight to other nodes is set to 0.9. The weight increases and decreases with the frequency of path usage.

The motion transformation matrix is responsible for issuing motion commands. For example, transfer cell AB is generated from location A to location B. This transfer cell is associated with the direction from A to B.

Each movement command is to arrive at the end position through a certain direction from a certain starting position. For example, to go from location A to location B, generate the command AB, which also includes the direction from A to B.

Autonomous robots, like humans, should also be designed with motivation. Motivation can be a task to be completed, looking for a power source to recharge when the battery is low, or returning to a fixed place after completing a task.

和所有方向角细胞的最大活动

其次，计算这两个最大活动的乘积，定义为

最后，得到路标方向角融合细胞的活动：The two information fusion of road signs and azimuth angles produces landmark azimuth fusion cells, which are intermediate variables for generating position cells, and the calculation method of its activity is divided into three steps. First get the maximum activity of the signpost cell

and the maximum activity of corner cells in all directions

Second, the product of these two maximal activities is computed, defined as

Finally, get the activity of the signpost direction angle fusion cell:

X ^Prd (t+1)＝[X ^Prd (t)+p] ⁺

When all waypoints have been explored, reset the activity of this cell.

A place cell is a variable that is only related to the location, and has nothing to do with the speed and direction of movement. This variable is defined as a "place cell", which corresponds to the biological place cell in function. Embodied as a variable in general-purpose DSP hardware. New locations that arise during exploration are encoded with new neurons. At a given location, multiple place cells are active and co-localized. The density of locations is related to the location of the environment where the robot is located. More landmarks are learned where the position-angle positions change rapidly, such as walls and doors. When the entire environment is learned, the environment is completely covered by place cells, each of which corresponds to a corresponding location. Place cells determine where the robot is located.

The signposts generated by the above process form "place cells" through neural network learning algorithms. Each location release is jointly determined by its own activity and landmark azimuth fusion cells (Fig. 3). If the robot is at the precise location expressed by place cells, its activity is greatest. When the robot moved away from this position, the activity of the place cells gradually decreased with the distance removed.

Each place cell is interconnected with all landmark azimuth fusion cells. Activity is determined by the scalar product of the activity vector of the landmark azimuth fusion cell with the corresponding connection weight vector. Therefore, the activity of place cells is determined by the learned partial view and the current partial view.

in $W_j=\underset{\mathrm{kl}}{\overset{N_{\mathrm{P.S.}}}{\mathrm{\Σ}}}W.$

The learning process of place cells follows the Hebbian law. When the robot is in a new environment, new neurons are automatically generated to encode new new positions. This happens automatically without external intervention. When previously learned place cell activity is within a given threshold, new neurons are automatically involved in encoding new places.

The above-mentioned signpost itself is a general concept. Here, the “local view around the feature point” is used to quantify the “signpost”. The variable representing the “signpost” is called the signpost cell. The local view around the feature point is used. The unit k represents that the landmark neuron k can be obtained by the following formula:

ΔW=I(t)□R

Where ΔW is the connection weight from pixel point i, j to the kth landmark, and the initial value is set to 0.I(t) The distance between the pixel point (coordinates i, j) at time t and the feature point. R is to represent whether this neuron participates in the encoding of this partial view, and the value of R is 0 or 1. 0 indicates that the connection weight is 0, and does not participate in the encoding of this partial view, and 1 indicates that it participates in the encoding of this partial view, and the connection weight is I(t).

The activity X ^Land (t) of the kth landmark cell is obtained by the following formula

${\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; Land</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N\; M</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&Delta;W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

N and M are respectively the number of pixels on the abscissa and ordinate in the partial view. f is defined as follows

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; Thr</span>}}{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Thr</span>}<span\; class="notranslate">\; ]</span>}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

where Thr is the recognition threshold. [x] ⁺ = x when x≥0, otherwise 0.

It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the protection scope of the appended claims of the present invention.

Claims (2)

1. A position cell bionic robot navigation algorithm, is characterized in that, comprises the following steps: A1. First, the robot’s wide-angle camera captures the panoramic photo, converts it into a gradient map, and then filters it with a Gaussian difference filter to detect feature points; perform logarithmic polar coordinate transformation on the local area near the feature points, and these feature points correspond to specific signposts; A2, for each road sign, obtain the angle relative to the true north direction, that is, the azimuth angle, and the true north direction is provided by the compass; the azimuth angle and the road sign jointly determine the position cell at the current moment; A3, the position cells at the current moment and the position cells at the previous moment jointly determine the transfer matrix; when the robot is exploring in an unknown environment, the transfer matrix is continuously formed, and the transfer matrix forms a cognitive map; A4, the transfer matrix and the cognitive map jointly determine the motion transformation matrix, which is responsible for issuing motion commands. 2. place cell bionic robot navigation algorithm according to claim 1, is characterized in that, described road sign adopts the partial view quantification around feature point, and the variable of characterizing road mark is called road sign cell, and the local view around feature point uses road sign Neuron k means that landmark neuron k can be obtained by the following formula: ΔW=I(t)□R Among them, ΔW is the connection weight from the pixel point i, j to the kth landmark, and the initial value is set to 0. I(t) the distance between the pixel point (coordinates i, j) at time t and the feature point; R is the distance representing the neuron Whether the element participates in the encoding of this partial view, R takes a value of 0 or 1, 0 indicates that the connection weight is 0, and does not participate in the encoding of this partial view, 1 indicates that it participates in the encoding of this partial view, and the connection weight is I(t) ; The activity X ^Land (t) of the kth landmark cell is obtained by the following formula ${\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; Land</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N\; M</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&Delta;W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span>$ N and M are respectively the number of pixels in the abscissa and ordinate in the partial view; the definition of f is as follows: $\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; Thr</span>}}{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Thr</span>}<span\; class="notranslate">\; ]</span>}^{<span\; class="notranslate">\; +</span>}$ where Thr is the recognition threshold. [x] ⁺ = x when x≥0, otherwise 0.

Images