aboutsummaryrefslogtreecommitdiff
path: root/doc/dui.md
blob: cbe99e9761db3de075fd8f6cf49cc0305bcf8546 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
# Introduction

\<inleiding iemand?\>

# Problem statement

The following is the original project description (translated to English).
References have been added in superscript that link to requirements set in
section \ref{specifications}.

> I would like to bring to market a vehicle that can drive
> independently\req{autonomous} from A to B. The vehicle must take into account
> traffic rules\req{traffic-rules}, road signs\req{signs}, traffic
> lights\req{traffic-lights}, etc. Research is being conducted using a small
> cart, the Pololu Zumo 32U4\req{zumo}, on which a camera module Nicla
> Vision\req{nicla} is mounted. The aim is to investigate the most appropriate
> method of recognizing the road, traffic signs and traffic lights. This should
> be demonstrated with a proof of concept. The cart does not need to drive
> fast\req{drspeed}, so the image processing\req{mvision} does not need to be
> very fast. Assume one frame per second (or faster)\req{imspeed}.

# Specifications

The following is a list of specifications derived from the original project
description in section \ref{problem-statement}.

\begin{enumerate}
  \item \label{req:autonomous}
    The vehicle is autonomous
  \item
    The vehicle can detect how its positioned and orientated relative to a road
  \item \label{req:traffic-rules}
    The vehicle conforms to the following set of traffic rules
  \begin{enumerate}
    \item Driving when not on a road is not allowed
    \item The vehicle can follow a road by steering itself accordingly
    \item Driving off the road is only allowed when necessary for the camera to
          keep seeing the road
  \end{enumerate}
  \item \label{req:traffic-lights}
  The vehicle handles traffic lights in the following way
  \begin{enumerate}
    \item Stop at a red traffic light
    \item Speed up at an orange traffic light
    \item Continue driving normally at a green traffic light
  \end{enumerate}
  \item \label{req:signs}
    The vehicle acts on traffic signs in the following way
  \begin{enumerate}
    \item Stop at a stop sign, and continue driving after a few seconds
    \item Turn left at a left sign
    \item Turn right at a right sign
    \item Slow down at a low speed limit sign
    \item Speed up to normal speed at a high speed limit sign
  \end{enumerate}
  \item \label{req:zumo}
    The vehicle used is a Pololu Zumo 32U4
  \item \label{req:nicla}
    The camera module used is an Arduino Nicla Vision
  \item \label{req:mvision}
    Computer vision / image processing is used as the only input
  \item \label{req:drspeed}
    There is no minimum speed the car has to drive at
  \item \label{req:imspeed}
    The image processing pipeline runs at 1 frame per second or higher
  \item
    The Zumo displays the name of the last detected sign on it's OLED display
\end{enumerate}

# Architecture

## Overview

![Architecture overview (level 0)
\label{fig:architecture-level-0}](../assets/architecture-level-0.pdf)

Figure \ref{fig:architecture-level-0} shows the hardware used in this project.
Both the Pololu Zumo 32U4 (referred to as just "Zumo"), and the Arduino Nicla
Vision ("Nicla") have additional sensors and/or outputs on-board, but are
unused.

## Distribution of features

Because creating a software architecture that does all machine vision-related
tasks on the Nicla, and all driving related tasks on the Zumo would create
significant overhead, and because the microcontroller on the Zumo is
significantly harder to debug than the Nicla, a monolithic architecture was
chosen. In this architecture, both the detection of 'traffic objects' and the
decisionmaking on how to handle each object is done on the Nicla board.

Figure \ref{fig:architecture-level-0} shows that a bidirectional communication
line exists between the Zumo and Nicla. This line is only used to send control
commands to the Zumo. Section \ref{niclazumo-communication-protocol} describes
which commands are sent over these lines.

## Nicla/Zumo communication protocol

The communication protocol used to control the Zumo from the Nicla uses UART to
send ranged numbers in a single byte. Table \ref{tab:protocol-ranges} shows
which number ranges correspond to which controls.

\begin{table}
\centering
\begin{tabular}{rl}
\toprule
\textbf{Description} & \textbf{Range (inclusive)}\\
\midrule
(unused) & \texttt{0x00}\\
Signs & \texttt{0x01} - \texttt{0x0f}\\
Speed & \texttt{0x10} - \texttt{0x1f}\\
Steering & \texttt{0x20} - \texttt{0xff}\\
\bottomrule
\end{tabular}
\caption{Protocol command ranges}
\label{tab:protocol-ranges}
\end{table}

### Signs

The Zumo stores the last sign received, and displays it's name on the OLED
display using the lookup table in table \ref{tab:protocol-signs}. The sign ID
is calculated by subtracting the start offset of the sign command range from
the command as shown in table \ref{tab:protocol-ranges}.

\begin{table}
\centering
\begin{tabular}{ll}
\toprule
\textbf{ID} & \textbf{Name}\\
\midrule
\texttt{0x00} & (clear sign)\\
\texttt{0x01} & Stop sign\\
\texttt{0x02} & Turn left\\
\texttt{0x03} & Turn right\\
\texttt{0x04} & Low speed limit\\
\texttt{0x05} & High speed limit\\
\texttt{0x06} & Traffic light (red)\\
\texttt{0x07} & Traffic light (orange)\\
\texttt{0x08} & Traffic light (green)\\
\bottomrule
\end{tabular}
\caption{Sign lookup table}
\label{tab:protocol-signs}
\end{table}

### Speed

The speed value ranges from \num{0} to \num{1}, and is converted from the
command using the following formula:

$$ v(n) = \frac{n - 16}{15} $$

### Steering

The steering value is similar to the speed value, but ranges from \num{-1}
(left) to \num{1} (right). The zumo has a built in "influence" value, which
limits the smallest radius the robot can turn at. The steering value is
converted using the following formula:

$$ s(n) = \frac{n - 32}{223}\cdot2-1 $$

## Zumo internal motor control functions

The Zumo robot receives a speed and steering value. Because the protocol has a
limited precision due to the low amount of data sent, the following formula is
used to control motor speeds $M_1$ and $M_2$ from steering value $s$ and speed
value $v$. The constant $C_1$ is used to globally limit the speed the robot can
drive at. $C_2$ represents the amount of influence the steering value has on
the corner radius, where \num{0} is no steering at all and \num{1} completely
turns of one motor when steering fully left or right:

$$ M_{1,2} = \frac{v(\pm s C_2 - C_2 + 2)}{2} C_1 $$

By default, $C_1 = \num{96}$ and $C_2 = \num{0.6}$

The Zumo firmware also smooths incoming values for $s$ and $v$ using a PID
controller. The default constants for the PID controller used are:

\begin{align*}
K_p &= -0.02\\
K_i &= +0.13\\
K_d &= -300.0
\end{align*}

# Research

## Detecting lines

The Zumo robot needs to drive in a road map-like environment where it needs to act like a car. With the nicla vision camera, there needs to be a way for detecting lines in every frame to make the Zumo robot ride between the lines. Read lines from an image there are different algorithms to make it work. We need to make sure that it works on the OpenMV program if we only choose this one. In this research, two techniques are researched: convolution-based and feature-based.

### Different line detection algorithms.

#### Hough Transform

This is a popular algorithm used to detect straight lines in an image. It works by transforming the image from Cartesian space to Hough space, where lines are represented as points. The algorithm then looks for clusters of points in Hough space, which correspond to lines in Cartesian space.

For more information about Hough Transform algorithms check the below links:

- [Wiki hough](https://en.wikipedia.org/wiki/Hough_transform  )
- [Science article](https://www.sciencedirect.com/topics/computer-science/hough-transforms)
- [OpenCV Hough](https://docs.opencv.org/3.4/d9/db0/tutorial_hough_lines.html)
- [OpenMV find_lines](https://docs.openmv.io/library/omv.image.html)

#### EDlines

EDLines, short for Edge Drawing Lines, is a feature-based algorithm that detects straight lines in an image by tracing along the edges of the image. It works by first extracting edges from the image, then building a graph where each edge is represented by a node. The algorithm then uses a greedy strategy to connect the nodes with high edge strength to form line segments. Finally, it merges line segments that are collinear and close to each other to form longer lines. This algorithm does not require a parameter search or optimization and is known for its robustness against noise and partial occlusion.

For more information about EDlines algorithms check the below links:

- [github library](https://github.com/CihanTopal/ED_Lib)
- [Science article](https://www.sciencedirect.com/science/article/abs/pii/S0167865511001772)
- [EDLINES: REAL-TIME LINE SEGMENT DETECTION BY EDGE DRAWING (ED)](https://projet.liris.cnrs.fr/imagine/pub/proceedings/ICIP-2011/papers/1569406487.pdf)
- [OpenCV EDlines doc](https://docs.opencv.org/3.4/d4/d8b/group__ximgproc__edge__drawing.html)

#### Line Segment Detector

LSD (Line Segment Detector) is an algorithm used for detecting line segments in an image. It works by analyzing the gradient information in the image and clustering nearby gradients that form a line segment. The algorithm first computes the gradient information for the image using the Gaussian filter. It then performs a series of operations, such as non-maximum suppression and thresholding, to obtain a binary edge map.

The line segments are detected by applying a series of geometric constraints to the edge map. These constraints include the minimum and maximum length of line segments, the minimum angle between line segments, and the maximum deviation of line segments from a straight line.

Once the line segments are detected, they are refined using a line merging algorithm combining nearby line segments into longer, more continuous lines. The resulting line segments and their endpoints are returned as the output of the algorithm.

For more information about Line Segment Detector algorithms check the below links:

- [LSD: a Line Segment Detector pdf](http://www.ipol.im/pub/art/2012/gjmr-lsd/article.pdf)
- [Working behind LSD](https://saiwa.ai/blog/line-segment-detection-2/)
- [OpenCV LSD doc](https://docs.opencv.org/3.4/db/d73/classcv_1_1LineSegmentDetector.html)
- [OpenMV find_line_segments](https://docs.openmv.io/library/omv.image.html)

#### Radon transform

Radon transform is another popular algorithm used for line detection. It works by computing the line integral of an image along different directions. The algorithm rotates the image at different angles and computes the sum of pixel intensities along each line in the image. The result is a two-dimensional matrix called the Radon transform. Peaks in this matrix correspond to the lines in the original image. The algorithm then applies some post-processing steps to identify and extract the lines from the Radon transform.

For more information about Radon transform algorithms check the below links:

- [Science article](https://www.sciencedirect.com/science/article/abs/pii/0031320396000155)
- [matlab Radon](https://stackoverflow.com/questions/35412573/radon-transform-line-detection)
- [Matlab elaboration Radon](https://www.kevinpolisano.com/Doctorat/doc-matlab-exemple/radon_lines_detection.html)
- [OpenCV Radon doc](https://docs.opencv.org/4.x/d5/d89/radon__transform_8hpp.html)

### Which algorithm is suitable for our project?

We have identified four different types of line detection algorithms that could potentially be used for our project. To decide on the best algorithm, we need to consider various factors such as accuracy, efficiency, and ease of use. While processing time is not a critical factor in our case, we need to ensure that the algorithm we choose meets our requirements and is accessible through the platform we are using, which is currently openMV but may change to openCV in the future. Therefore, our priority is to select an algorithm that is easy to implement, provides accurate results, and is compatible with our platform.

#### OpenMV

The only two algorithms that work with OpenMV are Hough Transform, the function find_lines, and Line Segment Detector, also known as find_line_segments. Both of these have their ups and downs and could be used for our project. find_lines has the most ups whereas find_line_segemtns has the most negative. As the result here below is decently optimized, it is first grayscaled, and then canny edge detection is done to it.

For the test are the straight lines pictures used with different lighting additionality the left lane represents a whitish line and the right lane is drawn with a more darker color. here below are the pictures used:

![picture 1](../RealTime_pictures/rtStraightLines.class/00000.jpg)

![picture 2](../RealTime_pictures/rtStraightLines.class/00018.jpg)

##### find_lines

The find_lines is a very fast function where you can handle straight lines and other lines with at least 45 FPS or more. Also, have a lot of control over the different types of parameters.

This is the outcome of picture 1:
![outcome_picture_1](../assets/hough_straightLines_Pic_0.png)

This is the outcome of picture 2:
![outcome_picture_2](../assets/hough_straightLines_Pic_1.png)

As you can see there isn't much of a difference between the two pictures.

##### find_line_segments

The find_line_segments is a very slow function where you can find segments from a line. This is a easier to use function because it only has two parameters but the frame rate drops significantly. Additionally, the size of the image to run the algorithm on needs to be smaller because of memory.

This is the outcome of picture 1:

![outcome_picture_1](../assets/LSD_straightLines_Pic_0.png)

This is the outcome of picture 2:

![outcome_picture_2](../assets/LSD_straightLines_Pic_1.png)

As you can see there is quite a lot of difference between them. This function needs more refinement but I couldn't find the sweet spot. Also, the right line in different pictures was always the problem, so there needs another solution for this function to work better.

#### OpenCV

All the above algorithms could be used with OpenCV, But the Radon transform needs more work than the others with the amount of information in the doc.

## Communication between the Nicla and Zumo

In order to make the Zumo robot both detect where it is on a road, and steer to
keep driving on said road, some sort of communication needs to exist between
the Nicla and Zumo. As mentioned earlier\footnote{dit is nog niet benoemd}, all
machine vision-related tasks will happen on the Nicla board. Because the Nicla
board is the first to know how much to steer the cart, it makes sense to have
it control the cart by giving the Nicla a 'steering wheel' of sorts.

This section tries to answer the question "What is the best protocol to use
over the existing UART connection between the Nicla and Zumo?". After a
brainstorm session, we came up with the following specifications for the
communication protocol:

1. **Low bandwidth**  
   Less data means more responsive steering
2. **As simple as possible**  
   The Nicla only needs to control speed and steering
3. **Easy to mock and test**  
   The cart should be able to be controlled using a mock driver and the Nicla's
   output should be testable (preferably using unit tests)
4. **Adaptive to noisy data**  
   The cart should gradually change speed and steering direction as to not slip
   or cause excessive motion blur for the camera module on the Nicla
5. **Adaptive to Nicla failure**  
   If the Nicla crashes or can't detect anything, it will stop sending control
   commands. In this case, the Zumo robot should slowly come to a halt.

Where possible, it's generally benificial to re-use existing code to save on
time. Existing code exists for a custom binary protocol and a text-based
command protocol. Both of these were designed without bandwidth or latency in
mind, and mostly focus on robustness in the case of temporary disconnects or
noise on the communication lines, so a new protocol needs to be made.

To address specification 1 and 2, the command length is fixed at 1 byte. This
means that UARTs built-in start/stop bit will take care of message start/end
detection, since most software interfaces for UART (including Arduino) string
multiple sequential messages together even if they're not part of the same UART
packet.

To mock messages from the Nicla to the Zumo robot, a simple USB serial to UART
cable can be used, along with a small C or Python program to convert
keyboard/mouse input into steering/speed commands. A small software layer can
be implemented on the Nicla to log the semantic meaning of the commands instead
of sending actual UART data when run in a unit test.

A PID controller can be used to smoothly interpolate between input
speed/steering values. This would also introduce some lag between when the
Nicla knows how much to steer, and when the Zumo actually steered the wanted
amount. Smoothing the speed/steering values does make it virtually impossible
for the Nicla to make it's own input data unusable because of motion blur, so
the lag needs to be handled in some other way as directly controlling speed
values without interpolation would lead to a garbage-in-garbage-out system. The
simplest solution to motion blur is limiting the maximum speed the Zumo robot
can drive at, which is the solution we're going to use as speed is not one of
the criteria of the complete system\footnote{Problem statement
(section \ref{problem-statement})}.

In the case the Nicla module crashes or fails to detect the road or roadsigns,
it will stop sending commands. If the Zumo robot would naively continue at it's
current speed, it could drive itself into nearby walls, shoes, pets, etc. To
make sure the robot doesn't get 'lost', it needs to slow down once it hasn't
received commands for some time. As mentioned in section \ref{TODO}, the Nicla
module is able to process at about 10 frames per second, so 2 seconds is a
reasonable time-out period.

\def\communicationConclusion{
The complete protocol consists of single byte commands. A byte can either
change the cart speed or steering direction, both will apply gradually. When no
commands have been received for more than 2 seconds, the Zumo robot will
gradually slow down until it is stopped. Exact specifications of commands are
provided in the protocol specification document\footnote{dit document bestaat
nog niet}.
}
\communicationConclusion

## Compiling and linking code for the Zumo

This section tries to answer the question "What are possible debugging options
for code running on the Zumo robot?".

Debugging running code can usually be done using an on-device debugger, and a
program that interfaces with the debugger and gcc on a computer. Because gcc
only gives valuable information when it has access to the executable file
running on the microcontroller, an attempt at making a makefile-based toolchain
was made.

Pololu provides C++ libraries for controlling various parts of the Zumo's
hardware. These libraries make use of functions and classes that are part of
the Arduino core library. The Arduino libraries are all open source, and can in
theory be compiled and linked using make, but this would take more effort than
it's worth, since the Zumo has very little responsibility in the end product.

\def\buildSystemConclusion{
Because making a custom build system would take too much time, and because the
Zumo robot's code is very simple, unit tests are used to debug the Zumo's code.
For compiling and uploading the full Zumo firmware, the Arduino IDE is used in
combination with the standard Pololu boards and Libraries.
}
\buildSystemConclusion

## Traffic Sign Detection (TSD)

The following chapters will look at the different methods used to detect signs in order to find a suitable solution for this project.

### Color based

The distinct color characteristics of traffic signs can attract drivers’ attention and can also provide important cues to design color based detection methods. In the past decades, a large amount of detection methods are designed to detect distinct traffic sign colors such as blue, red and yellow. These methods can be directly used for traffic sign detection, and can also be used for preliminary reduction of the search space,  followed by other detection methods

#### RGB

One can easily look at the RGB values to detect a certain color. Although the r,g and b values are heavily effected by different illuminations, therefore this isn't a reliable solution in variating lighting conditions.

An example implementation: 

```py
#Red
if(R >= ThR and G <= thG)

#Blue
if(B >= thB)

#Yellow
if((R + G) >= ThY)
```

It is possible to enhance the colors with maximum and minimum operations:

```py
fR(i) = max(0, min(iR  iG, iR  iB)/s),
fB(i) = max(0, min(iB  iR, iB  iG)/s),
fY(i) = max(0, min(iR  iB, iG  iB)/s).
```

This method can result in some issues on the blue channel (see source 1 page 86583 for more explanation). As a solution to this issue use the following formula for the blue channel instead:
```py
fB(i) = max((0, iB  iR)/s).
```

#### HSV

The HSV/HSI color space is more immune to the illumination challenges of RGB. The hue and saturation channels can be calculated using RGB, which increases the processing time.
The following pseudo code shows how to detect the red, blue and yellow colors in this space.

```python
#Red
if (H <= Th1 or 
    H >= Th2)

#Blue
if (H >= Th1 and 
    H <= Th2)

#Yellow

if (H >= Th1 and 
    H <= Th2 and 
    H <= Th3)
```

#### LAB

This color space is used for finding uncorrelated color components, the L\*a\*b\* space was used for detecting blue, red, yellow and green colors. 

#### Pixel classification

This method avoids the use of fixed thresholds that might need adjusting at times. In order to resolve this some authors tried to transfer the problem into pixel classification where a neural network classifies every pixel in the input image, the pixel classification algorithms are often slower than other color extraction methods.

#### results

\def\signDetectionColor{
The above described methods where also applied to a database in order to compare each method. This resulted in the conclusion that, using a normalized RGB space is giving the a mixture of most detections and least false-positve results. See source 1 page 86584 for the full report.
}
\signDetectionColor

### Shape based

Common standard shapes of traffic signs are triangle, circle, rectangle, and octagon. Shape characteristics used for shape detection include standard shapes, boundaries, texture, key points, etc.

#### Hough

See line detection hough.
<!-- TODO: not enough info found -->

#### Barnes *Fang et al* (fast radial symmetry)

Seemingly a 'simpler' method 
<!-- TODO: (currently no great references found) -->

#### Fourier
This method also deals with occlusion and morphing/rotating the shape flat again (when looking at it at an angle).

#### Key points detection
This, simply put, looks at the edges/corners in order to find an ROI.

### Color & Shape based
In some methods, the shape detection methods can be combined with color based methods to fulfill the traffic sign detection work.
For example, using a color based detection to find a ROI and using shape detection in that ROI to determine if it is a sign.

### Neural networks
<!-- TODO: -->

### results

\def\signDetectionShape{
Most shape based recognition methods are more complex than using a color based detection. But the method 'Fourier' seems the most useful as it can also deal with rotated and occluded objects. 
}
\signDetectionShape

## Traffic Sign Recognition (TSR)
After traffic sign detection or tracking, traffic sign recognition is performed to classify the detected traffic signs into correct classes.

![signs example](../assets/signs.png)

### Binary tree
The binary-tree-based classification method usually classify traffic signs according to the shapes and colors in a coarse-to-fine tree process.

### Support Vector Machine (SVM)
As a binary-classification method, SVM classifies traffic signs using one-vs-one or one-vs-others classification process.

## results

\def\signRecognition{
While making a binary tree is seemingly the most simple, yet effective solution. Using the 'Fourier' method is a bit more complex, it may be a better solution (this requires testing).
}
\signRecognition

# Conclusion

\communicationConclusion
\buildSystemConclusion
\signDetectionColor
\signDetectionShape
\signRecognition

## Sources:

1. [IEEE, Digital Object Identifier June 26, 2019 (pages 86578 - 86596)](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8746141)