| dc.description.abstract | Autonomous vehicles and driver-assistance features have become increasingly more common
over the past 5 years. With such an increase in autonomous-vehicle technology, algorithms that
identify and respond to the dynamic road environment are essential. Many current navigation
programs require that the travel environment be mapped numerous times before an autonomous
vehicle can travel with no assistance. However, the vehicle must be able to also navigate in cases
where the current path is unfamiliar. This thesis explores the use of computer-vision techniques
such as color masking, Hough transformations, and bird’s-eye perspective transformations for the
purpose of implementing a pure pursuit navigation algorithm on a previously unknown course.
The goal of this navigation program is to navigate the vehicle as closely as possible to the middle
of the lane while smoothly following the path trajectory. An additional goal of this project is to
implement a histogram of oriented gradients (HOG) detector for the identification of street signs
and adjust the speed of the vehicle accordingly. This detector should have a near 100% success
rate, and perform the detection more quickly than previously implemented object detectors.
The pure path-planning algorithm proved successful in maintaining the vehicle in the lane and
proved very adapt at following the slopes of turns. In the three turns on the track, the maximum
deviations from the center line were 9.14 cm (3.6 in), 2.3 cm (0.87 in), and 8 cm (3.15 in), which
is very good considering the sharpness of the turns. In the straightaways, the vehicle did not
perform as well, deviating 11.2 cm (4.4 in) and 15 cm (5.9 in) on the first and second straightaways,
respectively. This deviation is mostly due to the vehicle shallowing the steering angle too quickly
on the turn exit, leading to the vehicle not being centered in the lane heading down the straightaway.
However, the vehicle was able to perform all navigation with low latency, achieving 30 frames per
second (fps) compared to the 5 fps of previous lane tracking attempts. Additionally, the vehicle
achieved the main objective of remaining within the lane boundaries throughout the entirety of the
Autonomous vehicles and driver-assistance features have become increasingly more common
over the past 5 years. With such an increase in autonomous-vehicle technology, algorithms that
identify and respond to the dynamic road environment are essential. Many current navigation
programs require that the travel environment be mapped numerous times before an autonomous
vehicle can travel with no assistance. However, the vehicle must be able to also navigate in cases
where the current path is unfamiliar. This thesis explores the use of computer-vision techniques
such as color masking, Hough transformations, and bird’s-eye perspective transformations for the
purpose of implementing a pure pursuit navigation algorithm on a previously unknown course.
The goal of this navigation program is to navigate the vehicle as closely as possible to the middle
of the lane while smoothly following the path trajectory. An additional goal of this project is to
implement a histogram of oriented gradients (HOG) detector for the identification of street signs
and adjust the speed of the vehicle accordingly. This detector should have a near 100% success
rate, and perform the detection more quickly than previously implemented object detectors.
The pure path-planning algorithm proved successful in maintaining the vehicle in the lane and
proved very adept at following the slopes of turns. In the three turns on the track, the maximum
deviations from the center line were 9.14 cm (3.6 in), 2.3 cm (0.87 in), and 8 cm (3.15 in), which
is very good considering the sharpness of the turns. In the straightaways, the vehicle did not
perform as well, deviating 11.2 cm (4.4 in) and 15 cm (5.9 in) on the first and second straightaways,
respectively. This deviation is mostly due to the vehicle shallowing the steering angle too quickly
on the turn exit, leading to the vehicle not being centered in the lane heading down the straightaway.
However, the vehicle was able to perform all navigation with low latency, achieving 30 frames per
second (fps) compared to the 5 fps of previous lane tracking attempts. Additionally, the vehicle
achieved the main objective of remaining within the lane boundaries throughout the entirety of the
test. The HOG sign detector proved very successful, achieving a 100% success rate in detection of
both stop sign and speed limit signs. The detections occurred in an average 0.716 and 1.07 s for the
stop sign and speed limit sign respectively, both faster than the 1.1 s of previous object detection
attempts.
The results of this work demonstrate the potential application of the path-planning system as
a backup to the current autonomous-vehicle navigation systems in situations where the route has
not been previously mapped manually by a human driver. The success of the HOG detector shows
great promise in being used in applications where the object being detected has a consistent shape,
such as the numerous warning signs and route signs on highways. | en |
