Funktionen

SLAM with Kalman-Filtered Odometry

Background and Problem Statement

Autonomous driving is currently a focus of attention in the automotive industry. Formula Student Germany (FSG), a competition for young engineers, is one of numerous yearly events with a specific event category for driverless vehicles. To succeed in this event with a productive, competitive vehicle, a team of dedicated students takes sole responsibility across all project stages, from first concepts to the final construction of the vehicle and finally its monitoring on a race track. The major responsibility of computer scientists in these teams is to conceptualize and integrate software control loops which are essential to operate an autonomous vehicle with respect to constraints and requirements of a racing scenario.

The team municHMotorsport is competing in the driverless category of the FSG events and remodeled its competing electric car from the previous year to fit the new needs of the autonomous events.

A new software sensor substystem has to be introduced in the vehicle control systemt to exploit new environmental and odometry sensors to support automated steering in maintaining a stable and efficient track. In particular, sensor outputs from a deep learning object detector at 15Hz soft real-time and a 1kHz real-time odometry sensor unit are to be combined to produce real-time sensor output. Using Kalman filters, reliability and precision of the combined sensor ouput can be significantly improved compared to the two original sensors and with constant, deterministic overhead.

This thesis should provide a systematic model of the SLAM (Simultaneous Localisation and Mapping) software subsystem for formal evaluation. It is then realized as a reference implementation and integrated in the existing sensoric pipeline of the productive vehicle control loop.

In this, a major priority is the formulation (mathematical and as system models) and evaluation of Kalman filter algorithms in the localization and mapping control loop component with respect to the racing event regulations specified by the FSG. Ideally, the resulting SLAM component comprises a computationally inexpensive framework with constant runtime complexity and exhibits a modular structure to alleviate modifications like integration of additional sensor systens.

Provided Sytem

• A fully functional, manually operated electric car (PWe7.18)
• Sensor components to be combined are integrated and functional:
  • Deep-Learning object detection system which localizes cones along the track with varying accuracy
  • Odometry sensors, tuned and accessible via ECU connection

Contributions

• Analysis of the properties of localization and mapping algorithms wirh respect to FSG specification
• A concept (control theory system model) to fuse the given environmental sensors and odometry sensors which have different soft/hard real-time clock frequencies
• Design of a Kalman filter to combine sensor outputs into unified, improved output.
• Discussing and evaluating approaches to make the interface compatible with additional environmental sensors
• A Reference implementation of the SLAM subsystem and its integration in a productive vehicle.
• Profiling and evaluation of the reference implementation in simulated scenarios using recorded real-world vehicle sensor data
• If feasible within the time frame, the reference immplementation will also be evaluated on track, in the competitions in FSItaly and FSGermany with the car PWd2.18

Requirements

• Some experience in control theory, at least interpreting and designing models of control systems
• Experience with Kalman filters and/or their mathematical foundations
• Productive experience in C/C++, ideally for real-time requirements

Related Work

• Extended Kalman Filter
• MonoSLAM
• Introduction into Gaussian Mixture Models

Duration:

• Masterarbeit: 6 Monate

Number of Students: 1

Supervisor:

• Tobias Fuchs