ROS BASED SELF-DRIVING RC CAR WITH TENSORRT INTEGRATION IN REDUCING BEHAVIORAL CLONING LATENCY AND OBJECT DETECTION FOR COLLISION PREVENTION
Renzo Llenard Monteadora
Explore the full thesis here»
A study was conducted to demonstrate a way to increase the performance (latency) of the behavioral cloning model and be able to accommodate additional computer vision task (object detection) even with a small-scale prototype (RC Car) using Jetson Nano. Microprocessor overclocking and utilization of TensorRT were done in the study to increase the performance of the behavioral cloning model in the prototype. Significant reductions in latency were found after doing such procedures. In overclocking, the performance of the behavioral cloning model was increased by 25% depending on the complexity of its CNN architecture, then after the integration of TensorRT, an approximate double (2x) of performance was added. Through the use of the Robot Operating System (ROS), an object detection was integrated into the system to enable collision prevention feature on the prototype, which lessens the performance behavioral cloning model by almost half (½). The study covers every step of the process in completing the autonomous self-driving RC car: from the assembly of the RC Car model; creation and training of the behavioral cloning model; robot operating system (ROS) structure, including object detection integration; overclocking and utilization of TensorRT; until the testing and evaluation of the prototype on a racetrack.
Below are the demo videos of the prototype using Behavioral Cloning Only and with additional Object Detection Model.
Two convolutional neural networks that were used in the prototype are introduced. It contains an overview regarding how it was constructed in the program using the TensorFlow framework, which can be found in the model.py script.
In the study, Robot Operating System (ROS) were used as a middleware platform of the prototype. The figure below shows the ROS Map of the system.
Distributed under the MIT License. See LICENSE.txt for more information.
Renzo Monteadora
- Linkedin: Renzo Llenard Monteadora
- Twitter: @renzo_llenard
- Email: renzk246@gmail.com
- Ace Virgil D. Villaruz, MSEE - Adviser