.png){kind=icon}
.png){kind=icon}
Return
Real-time edge AI algorithm for road defects detection
.png){kind=icon}
PRE
AI
DOI：10.1080/14680629.2025.2460472.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Vision-based automated road defects detection schemes are significant for highway maintenance and road condition grade assessment. This paper was dedicated to achieving a refined trade-off between accuracy and efficiency through neural network optimisation. Specifically, we proposed a novel lightweight model, RT-RDD (Real-Time Road Defects Detection), which integrated low-level extraction and high-level semantic feature extraction for road defects detection. This was achieved by designing a more advantageous backbone structure and a lighter neck structure. Additionally, key improvements include optimising the feature extraction strategy and evaluation. The RT-RDD model surpasses existing models in terms of mean Average Precision (mAP), with only a marginal reduction after quantisation. Compared to foundation model, our algorithm improves mAP (0.5:0.95) by 4.2%. In addition, we propose a quantisation and compression strategy that effectively reduces the overall model size by approximately one-third. Notably, this reduction in size only results in a minor decrease of 1.3% in mAP (0.5:0.95). Testing on NVIDIA Jetson Xavier NX Development Board Kit demonstrates the model achieves a detection speed of 58 frames per second (FPS) with enhanced precision. These improvements make it ideal for deployment in practical road maintenance scenarios.
Keywords:
Road defects detection
edge devices
Lightweight Neural Network
quantisation and compression strategy
NVIDIA Jetson Xavier NX Deployment Board Kit
Journal
IF:
3
Papers:
2.5K
Citations:
6.9K