A Comparison of FPGA and GPGPU Designs for Bayesian Occupancy Filters

Area of Science:

• Computer Science
• Robotics
• Electrical Engineering

Background:

• Grid-based perception is crucial for automotive systems, fusing sensor data for environmental awareness.
• High computational demands of these systems pose challenges for embedded automotive applications.

Purpose of the Study:

• To evaluate the performance of novel computing architectures for automotive perception algorithms.
• To compare optimized Bayesian Occupancy Filter implementations on General Purpose Graphics Processing Units (GPGPU) and Field-Programmable Gate Arrays (FPGA).

Main Methods:

• Two specialized Bayesian Occupancy Filter designs were developed: one for GPGPU and one for FPGA.
• Implementations were tested and compared using data from a realistic simulator and a real automated vehicle.

Main Results:

• The study analyzed development effort, accuracy, and performance metrics for both GPGPU and FPGA implementations.
• Performance and accuracy trade-offs between the two architectures were identified in real-world automotive scenarios.

Conclusions:

• New computing architectures can address the high computational requirements of automotive perception systems.
• The choice between GPGPU and FPGA depends on specific project needs regarding development effort, accuracy, and performance.