Real-time detection of pedestrians in night-time conditions using a vehicle mounted infrared camera
Hurney, Patrick
Hurney, Patrick
Loading...
Identifiers
http://hdl.handle.net/10379/6374
https://doi.org/10.13025/17124
https://doi.org/10.13025/17124
Repository DOI
Publication Date
2016-09-30
Type
Thesis
Downloads
Citation
Abstract
Current statistics show that a significant number of road fatalities occur during night-time hours despite a smaller number of vehicles on the road. This number could be significantly reduced with the use of systems that automatically detect Vulnerable Road Users (VRUs) and alert the driver of their presence. This thesis presents an efficient embedded Advanced Driver Assistance System (ADAS) to detect pedestrians with a far-infrared sensor in real-time. The ADAS proposed in this thesis is implemented on a low power Intel Atom Central Processing Unit (CPU) and an Altera Arria GX II Field Programmable Gate Array (FPGA). The CPU and FPGA communicate over a PCI-express 2.0 bus using a Direct Memory Access engine on the FPGA. The pedestrian detection algorithm was partitioned between the FPGA and CPU to ensure real-time performance. Far-infrared images are first acquired with a microbolometer at a rate of 25fps, and a morphological closing operation is applied to remove distortion on the pedestrian's torso. The frames are sent to the FPGA to be processed using dedicated hardware. Regions of Interest (ROI) that could potentially contain pedestrians are isolated from the background using hardware accelerated Seeded Region Growing (SRG). The isolated ROI are classified using a Support Vector Machine (SVM) on the CPU. Histogram of Oriented Gradient (HOG) features and Local Binary Pattern (LBP) features are extracted from the ROI. These features are then concatenated to form a HOG-LBP feature vector and are passed to a classifier that determines if the ROI contains a pedestrian or non-pedestrian object. Successfully classified pedestrians are tracked between frames using a Kalman filter. The system runs in real-time at a rate of 25fps, the frame rate of the micro bolometer. The hardware accelerated SRG method obtained a 97.93% reduction in execution time compared to the software implementation than on the CPU alone. Detection rates of 98% have been achieved with HOG-LBP features on a database of 2,000 individual frames and video containing 15,000 frames. The total power consumption of the system is 3.12W. This results in a highly accurate, low power, low cost system suitable for the automotive sector.
Funder
Publisher
Publisher DOI
Rights
Attribution-NonCommercial-NoDerivs 3.0 Ireland