Low-Cost Blind Spot Detection System Based on Lite Object Detection Algorithm and Limited Resources Hardware

Blind spot detection systems have become integral components of modern vehicle safety systems, aiming to mitigate the risks associated with lane-changing maneuvers. However, existing blind spot detection solutions often rely on expensive external sensors or complex infrastructure such as Radar and LIDAR, limiting their widespread adoption, particularly in budget-conscious markets. In this work, the combination of computer vision techniques and edge computing to detect vehicles is used in the blind spot region in real time. A lightweight convolutional neural network (CNN) model is deployed on an embedded device, such as a microcontroller, to process video streams from onboard cameras. The CNN model analyzes the captured frames, identifying vehicles near the host vehicle. This study compares the performance of several algorithms, including SSD FPN-Lite, YOLO, and FOMO, using a dataset of automated rickshaws, with images captured at various angles and dimensions. It has been shown that the FOMO algorithm is the best option for working on microcontrollers. An Arduino was used to connect to a camera that captures images in real time, recognizes the object, and identifies it with a center point. The best possible testing accuracy for this work using the collected dataset was $88.24 \%$.