Embedded Night-Vision System for Pedestrian Detection

Abstract

Through this study, we show that contemporary computer vision algorithms and readily available hardware may be used to create a low-cost, mobile pedestrian detection system that can function well in challenging lighting circumstances, such as fog or dusk. Our suggested solution uses an ODROID XU4 microcomputer running Ubuntu MATE in conjunction with a thermal imaging camera. A compact yet potent single-board computer, the ODROID XU4 was chosen for its ideal balance of portability, processing power, and affordability.
This technology relies heavily on thermal imaging, which is resistant to the limits of visible light since it can detect infrared radiation that things release. This makes it possible to detect pedestrians reliably in glare or darkness. We constructed a cascade object detector specifically designed for thermal images as part of the detection method. In order to identify human silhouettes using their distinct thermal signatures, a classifier must be trained. It operates by filtering portions of a picture one after the other, rapidly eliminating the least likely locations, and focusing processing on areas with a high probability of containing pedestrians, which makes it a computationally effective option for resource-constrained embedded systems. We used a deep learning-based method, a convolutional neural network (CNN) tuned for object detection in thermal images, to evaluate system performance in comparison to the cascade object detector. Although CNNs often offer improved detection accuracy, their significant computational resource requirements can be a barrier for portable, cost-conscious applications. Our findings demonstrate that a properly calibrated cascade detector can achieve similar performance in a variety of situations while using a lot less power. Advanced driver-assistance features, including night-vision pedestrian recognition, don't have to be limited to high-end cars, as this work shows. Effective and accessible solutions that can improve road safety for a larger variety of drivers can be developed by utilizing thermal imaging, effective algorithms, and reasonably priced computing systems. Intelligent vision-based transportation solutions are made more accessible with this method.