Illumination and Temperature-Aware Multispectral Networks for Edge-Computing-Enabled Pedestrian Detection

Accurate and efficient pedestrian detection is crucial for the intelligent transportation system regarding safety mobility, e.g., Advanced Driver Assistance Systems, smart crosswalk systems. Among all methods, vision-based method demonstrated to be most effective in previous studies. However, existing algorithms still have two limitations that restrict their implementations, those being real-time performance as well resistance impacts of environmental factors, low illumination conditions. To address these issues, this study proposes a lightweight Illumination Temperature-aware Multispectral Network (IT-MN) accurate detection. The proposed IT-MN an one-stage detector. For accommodating factors enhancing sensing accuracy, thermal image data fused by with visual images enrich useful information when quality limited. In addition, innovative late fusion strategy also developed optimize performance. make model implementable edge computing, quantization applied reduce size 75% while shortening inference time significantly. algorithm evaluated comparing it selected state-of-the-art using public dataset collected in-vehicle cameras. results show achieves miss rate at 14.19% 0.03 seconds per pair on GPU. Besides, quantized 0.21 device, demonstrating potentiality deploying devices highly algorithm.