【摘要】 当前，许多用于检测头盔佩戴情况的算法在低算力嵌入式设备上无法满足实时检测的需求，从而限制了头盔佩戴检测技术的广泛应用。针对这一难题，本文提出了一种针对Jetson Nano开发板的TensorRT优化部署方法。首先，采用Int8量化、层间融合和张量融合等技术提升算法性能和加速推理速度；然后，利用TensorRT的自动化校准过程，使算法性能损失最小，解决使用Int8导致信息丢失的问题。实验表明，将头盔佩戴检测算法模型部署到Jetson Nano嵌入式设备中后，mAP@0.5达到98.63%，推理总耗时从320.52 ms减少到64.11 ms，减少了80%。这一改进有效地解决了算法在低算力嵌入式设备下部署推理的实时性不足的问题，为头盔佩戴检测技术的推广应用提供了新思路。更多还原

【Abstract】 Currently, many algorithms used to detect helmet wearing conditions cannot meet the real-time detection requirements on low computational power embedded devices, thereby limiting the widespread application of helmet wear detection technology. To address this challenge, a TensorRT optimization and deployment method for the Jetson Nano development board has been proposed. Firstly, algorithm performance and accelerate inference speed are improved by using the techniques such as Int8 quantization, inter-layer fusion,and tensor fusion. Then, the loss of algorithm performance is minimized by using TensorRT’s automated calibration process, addressing the problem of information loss caused by the use of Int8. The experiments show that after deploying the helmet wear detection algorithm model into the Jetson Nano embedded device, the mAP@0.5 of the algorithm achieves to 98.63%, and the total inference time reduces from 320.52 ms to 64.11 ms, which is decreased by 80%. This improvement effectively solves the problem of insufficient real-time deployment inference of the algorithm on low computational power embedded devices, giving a new method for the widespread application of helmet wearing detection technology.更多还原