Despite a set back caused by the fact that last Friday my personal computer stopped working, this week was mostly focussed on writing the results chapter of the masters dissertation and performing some experiments to gather result regarding the developed systems.
These experiments include the study of the accuracy with which the roll and pitch are translated in representation of the laser scans. For this an hydraulic jack was placed on the rear of the atlas car, allowing to slowly increase the car’s pitch, and the car was placed in front of a flat surface, from this experiment the pitch and roll where gathered along with the variable height of a laser scan point that intersected the plane, the setup can be seen in the figure bellow along with a graph detailing the variation and a short video to help visualize the results.
Other tests have been made regarding the accuracy for positioning of the atlas car, using the mapviz application to visualize a circuit travelled by the atlas car. Regarding the road reconstruction module some results have also been gathered and evaluated, as well as some troubleshooting and subsequent corrections to the code which was not correctly representing the points on the ground plane since it lacked the parameter for the orientation of the lidar sensor.
That said as of this moment only half of the results chapter as been written, only missing the results and analysis of the experiments performed in the simulator and the actual real-world tests of the algorithm.
These past two weeks were almost completely spent writing the Chapters 2, 3 and 4 of the dissertation.
Despite this in this week, with the help of the colleague Pedro B. Nova, responsible for the global navigation of the AtlasCar 2, the road edge detection package was integrated with the mapviz application used by Pedro. The processe involve a slight tweak to the frame tree so as to receive the position from the node published by global navigation node.
The integration allows to overlap the road edge point cloud and the satellite map provided by the map viz application, and provides a way do qualitatively evaluate the accuracy of the boundary detection method. The images bellow show some of the obtained results.
From these results we can see that the detection is pretty accurate and the resulting point cloud almost always overlaps the borders represented on the image. It even detects the two platforms on the same side (to the right), near Complexo Pedagógico.
There are still some outlier points that need to be removed but overall the quality of the road limits detection is very good.
Week 9 was primarily spent on rebuilding the simulator, correcting some mistakes previously made, in order to more accurately measure density according to the method previously discussed. Following this an analysis of the case study was performed considering that the sensor was static relative to the road. Several configurations of side-walk were used as well as several sensor configurations, in term of its height and angle, though as was expected the results of little consequence given that this method heavily relies on accumulated data.
Also in this week a “skeleton” of the thesis was written and presented to the thesis advisor, so that the dissertation structure could be discussed.
Week 10 was mostly focussed on writing the first chapter of the master’s dissertation and spending the time to research the state of the art for road boundary detection.
In regards to the practical aspect of the thesis, the simulator was adapted to be able to quantify the measured density in the case of a moving vehicle and some results were obtained.
The case study involved several configurations of vehicle velocity, sensor placement, curb position and angle, and radius of the filtering circle. Some of the results are presented below in the form of a graph. These result where are based on a 12 m displacement of the car with the filtering point at 14m from the initial sensor position, so in these graphs the X represents the displacement of the sensor (vehicle).
Since a lot of parameters were analysed here follows some of the conclusions drawn from them:
- Vehicle velocity:
- For smaller vehicle velocities the final density is much greater, but more time is required to obtain the same density levels when compared the greater speeds.
- Filtering parameters:
- For the tested radii, 0.2 m seems to be the one that produces the most density.
- The density on the ground plane is much less than on the vertical planes, and the radius of the filtering circle doesn’t have any major influence on the density results.
- Side-walk configurations:
- For the various side-walk configurations (angle to the ground plane \____/ ) despite the similar curve behaviour, the left side of the road (farthest from the sensor) tends to accumulate points faster and ends up with a greater density in almost every angle (except those closer to ground level). Also in these configurations the more perpendicular to the ground the greater the density.
- For side-walk configurations which rotate inward along Z ( / . \ ) the left side produced much greater densities but the behaviour remained the same in both sides, with the ideal angle for the most density being 45º.
- Sensor placement:
- The ideal height for most density seems to be between 0.35-0.4 m.
- The ideal angle for the sensor appears to be between 0.6º and 1.2º.
Despite the limited cases studied (most of them with the accumulation area in the same position) these graphs show results coherent with what was to be expected, and in some cases such as the sensor placement show that the configuration in the AtlasCar2 seems to be the ideal for the most density, which is really interesting.