Automated RC car for the CaroloCup

The goal is the design, development, and construction of an automated RC car in scale 1:10 to participate at the annual CaroloCup in Braunschweig, Germany. This project is in development by the student team "Team Galaxis" of RWTH Aachen.

Project State

Public Project

Licences

Software Licence: Project has no software
Hardware Licence: Project has no hardware

Project Tags

Admins

TeamGalaxis

Does this project pique your interest?

Login or register to join or follow this project.

Introduction

The Carolo-Cup is an annual student-competition taking place in Braunschweig, Germany. The challenge is to develop, produce and demonstrate a cost- and energy-efficient concept of an automated RC car in scale 1:10. The competition is split into three disciplins:

Simple Track:
The car has to follow a circuit which only contains crossings and parallel respectively perpendicular parking spaces. Beside staying within its lane, the car does not have to follow any traffic regulations. During the 2 minute drive, the car has to cover as much distance as possible and perform two parking maneuvers. Each mistake, i.e. leaving its lane or hitting an object during the parking maneuver, is penalized by loosing a certain amount of driven meters. In case car cannot continue on its own, one team member is allowed to intervene using a remote-control to steer it back onto the track. This will be indicated by a blue LED on top of the car and also results in a penalty.

Advanced Track:
In this discipline, the track contains additional elements and therefore more complex scenarios. Those elements are speed limits, static and dynamic obstacles, crosswalks with pedestrians, extended regulations at intersections (stop, give-way lines,  right-of-way, required direction to cross), no-passing zones. Every regulation will be indicated by a corresponding traffic sign and a marking on the road surface. Again, penalties are assigned for every violation or intervention.

Static Discipline:
Additionally to the demonstration of the car, the team has to give a presentation about the concept and the development process. This discipline will be rated by a jury composed of people working in the automotive industry.

Further information about the competition can be found here. This year's competition can be watched here (Startpoint set to our run.).

Hardware

Main-Computer: NVidia Jetson Tx2
Carrier-Board: Auvidia J120-IMU
ECU: STM32f407


Camera: Basler daA1600-60uc (sponsored)
Lens: Lensagon BM2920S118
CPL: Hoya Revo cir-pl 52mm

Other sensors:

  • 9-DOF sensor: Bosch Sensortec BNO055
  • 2x Time-of-flight distance: ST VL53L0X
  • Rotary encoder: austriamicrosystems AS5045


Battery: Turnigy multistart LiPO  3000mAh 3S 11.4V
Actuators:

  • Brushless motor: Team Orion Vortex  VST2 Pro LW 540
    • ESC: Team Orion Vortex R10 SC
  • Servos for front and rear axis: Savöx SB-2252MG

Remote-Control: DJI PVT581 5.8GHz

Software

The vision tasks can be separated into lane detection, traffic sign/ground marking detection and obstacle/pedestrian detection.

Lane detection:
The task here is to detect a two-lane street which limited by white lines on a black ground and output it in some usable form-in our case as third-degree polynomials. There are many difficulties for which a robust approach is necessary. Some marking may be missing, the center line can consist of different patterns (drawn through, dashed or combinations) and different lane widths. Additionally, there are many distracting markings and objects on the road.

We based our approach on He et al [1] which consists of three phases:

  • Preprocessing
    Find candidates for lane markings and extract them to feed the CNN
  • Siamese CNN
    Adds the probability of being a real lane marking
  • Optimization
    Create valid sets of lane markings and select the best combination using a loss function. For example sets consist of at most the lane markings, minimum and maximum lane widths are considered and the curvature is compared.

[1] He, Rui Ai, Yang Yan and Xianpeng Lang - Accurate and Robust Lane Detection based on Dual-View Convolutional
Neutral Network

Remaining Detections:
Unfortunately, all other detections did not get a chance to be tested in the competition. Therefore, we omit the descriptions by now and might add them later.

 


 

 

 

Comments

Your comments, please!

Want to comment this ... Show more