Real-Time Pedestrian Detection System Using OpenCL-Based FPGA Acceleration

Abstract

Pedestrian Detection System is a system that detects the presence of pedestrians in the vicinity of a vehicle. When a pedestrian is detected in the region of interest (ROI), it gives the driver an appropriate warning and/or takes appropriate braking action such as slowing down/stopping, swerving if avoidance is possible or deploying external airbags if collision is unavoidable, etc.

In a vision-based setup, the major issue is the requirement of high computational speed in order to complete the image processing as well as classification in real time. The processing and classification should result in detection of pedestrians fast enough that there is time for appropriate actions to be taken.

In order to be fully real time, an ADAS should have an upper bound on the time it will take to give results, including all corner cases. The system should also be feasible to be put on board a vehicle and be power and performance efficient.

This research proposes a new architecture that implements a Pedestrian Detection System in a heterogeneous computing environment and obtain results in constant time and can function in real-time. The proposed solution uses FPGA as a co-processor and the architecture is implemented in the heterogeneous programming framework OpenCL to yield the result in real-time.