Zamiast przeglądać losowe przykłady z githuba wysyłane przez komuhera, których później się wypiera, przeglądnąłem kilka artykułów o mierzeniu prędkości. Tak jak pisałem zawsze pojawia się krok ze skalowaniem do miar ze świata rzeczywistego. Może być to cokolwiek, szerokość pasa ruchu, rozmiary samochodu, prędkość kamery itd
Nie no masz racje reasearch w ktorym bralo udzial 20 topowych uniwersytow vs Hindus z medium oraz odpowiedzi sprzed 5 lat XD (pro tip ML wtedy raczkowal i to ledwo co)
I jedyny sensowny link jaki podeslales czyli jakis papier od razu mówi o tym ze algorytm nie korzysta z zadnych informacji dotyczacych miar odleglosci wpisanych na sztywno ani ustawien kamery tylko i wylacznie ze zdjec.
Our main contribution is that we directly learn velocity from images. No extra
calculations have to be performed to determine the velocity of the vehicle. Additionally,
we will learn velocity from the vehicle’s point of view (ego-vehicle speed). Traditional
optical flow approaches are static cameras that track movement in the image. The same
approach is used to track vehicles on the road. We place a camera on a vehicle and
use those images to learn velocity. This will cause each pixel in the image to move,
leaving our model to learn velocity when each pixel in the image has moved. Though
much research was conducted into measuring velocity, this has led to new sensors such
as accelerometers that can only measure velocity. With our approach, a camera can
be used to determine the velocity of the vehicle. The images can be used to determine
the velocity of the vehicle, but image analysis techniques can be used to determine the
environment of the vehicle and more. To our knowledge, there has been no research into
learning velocity from two consecutive images before.
I znowu mylisz pojecia -> algorytm vs dane treningowe co z tego ze zebrane dane sa z predkosci GPS'a to nie znaczy ze budujac algorytm musisz mu podawac dane z GPSa
5.2.1 Data set <-------
The data was created by equipping a car with static high-resolution colour and grayscale cameras and driving through streets of Karlsruhe, Germany. The ground truth was collected by using a Velodyne laser and
29 GPS, where the speed of the vehicle is can be calculated to meters per second (m/s).
https://esc.fnwi.uva.nl/thesis/centraal/files/f366183560.pdf
Dalej kolejny podpunkt
5.1 Experiment 1: Measuring Time
We will use TimeNet to predict the time that has passed between frames. The model
must learn what time looks like, by analysing the size of the displacement.
Given a stationary camera at the side of the road that records images at a fixed
frame rate (FPS) of a vehicle driving by at a fixed speed. Between frames f0 and f1,
the car will have moved over a certain distance. Since the distance to the vehicle and
the speed of the vehicle are fixed, the model will be able to learn how much time has
passed. If we keep the velocity of the vehicle the same, but increase the time between
f0 and f1, the vehicle will have travelled a larger distance. If TimeNet can learn the
relation between the magnitude of these displacements, it will have learned how much
time has passed between the frames.
The observed object will have to move at the same pace and distance across the receptive field of the camera. If the object moves closer or changes pace, the displacements
across the image will change. This will make it difficult for TimeNet to learn time from
image pairs
I kolejny
5.2 Experiment 2: Predicting Velocity
This experiment entails the core of our research; for the model to learn velocity. We will
use SpeedNet to do this. SpeedNet must learn that the magnitude of each displacement
can be different over the same time span and learn that this indicates speed. If the
model is successful to map the size of a displacement to a speed, the model has learned
what speed looks like.
Given a stationary camera mounted on a car that records images at a fixed FPS, the
camera will capture images while the car drives around. By looking at all moving parts
in the image, SpeedNet must determine how fast the car is driving. This is a different
problem to learning time, since each pixel in the image will move and there are no fixed
speeds. Additionally, there is other traffic, pedestrians and buildings that could make it
difficult for the model to generalise well. However, if the model is successful in building
an internal representation of each speed it will be able to predict speed.
I na koniec wyniki
Average error To gain insight into our models’ performance, we track their average
prediction error, which is displayed in Table 3. It is shown that SpeedNetSimple has
the smallest average prediction error over the entire test set, whereas LassoCorr has the
largest average prediction error. Between the Correlation and Simple models, the Simple
models outperform the Correlation models.
Model Average error
LassoCorr 3.75
LassoSimple 2.74
SpeedNetCorr 2.39
SpeedNetSimple 2.05
Table 3: Average error per measured velocity in m/s, per model.
Jest to papier powstajacy w roku 2016 wiec 3 lata temu dlatego ja podsylalem linki do rozwiazan sprzed roku ale to tez wystarczy,
Dziekuje do widzenia XD
Przytocze jeszcze tylko problem o ktorym jest dyskusja
Czy ... da się wytrenować siatkę sztucznej inteligencji aby na podstawie obrazu z kamery umieszczonej nad gruntem (nie zawierającym żadnych wystandaryzowanych znaczników czy miar odległości) potrafiła określać szybkość poruszania się pojazdu (w przedziale 0-30 km/h)?