A Cost-Effective Blind Spot Detection System with High Recall Rate using Deep Learning and a Comparative Analysis of Implementation Hardware Platforms

Blind Spot Detection (BSD) is critical for enhancing vehicle safety and is an integral component of many Advanced Driver Assistance Systems (ADAS).In this work, we propose a novel, cost-effective BSD solution utilizing a deep learning approach with the vehicle's existing rearview camera.While previous works such as Histogram of Oriented Gradients (HOG) combined with Support Vector Machine (SVM) have been used for BSD, these approaches often struggle to maintain high accuracy, particularly in diverse real-world environments.To address these limitations, we employed ResNet-18, a deep learning network, to improve both recall and precision in detecting vehicles in the blind spot.Our approach was evaluated across multiple hardware platforms, including CPU, GPU, and Field Programmable Gate Arrays (FPGA).We conducted comparative analyses in terms of detection performance, processing speed, energy consumption, and cost efficiency.The results showed that our deep learning-based BSD system achieved a 100% recall rate on all hardware platforms, ensuring no critical events were missed, thereby greatly enhancing safety.Among the platforms, FPGA demonstrated superior energy efficiency.Over time, FPGA emerged as the most cost-effective platform due to its low operational costs.This work contributes to the development of more reliable and efficient BSD systems by leveraging deep learning and identifying the optimal hardware platforms for real-time vehicle detection in ADAS applications.