基于Jetson TX2的路面裂缝检测系统设计

Design of Jetson TX2-based Pavement Crack Detection System

针对目前采用深度学习框架的路面裂缝检测方法存在落地应用难、成本高与效率低等问题,设计了基于Jetson TX2的路面裂缝检测系统。通过YOLOv5网络识别路面裂缝,使用U-Net网络对裂缝目标进行分割,并根据分割结果进行路面健康评价;其次,利用TensorRT方法优化深度学习模型,提高推理速度;最后,结合DeepStream框架设计路面视频流分析系统并部署到Jetson TX2嵌入式平台。实验结果表明:路面裂缝目标检测模型对横向、纵向和网状裂缝3种路面常见路面裂缝的检测精度均达到了90%以上,且模型优化后的推理速度为30.7ms/帧,速率提升35.1%;最后经过验证,Jetson TX2嵌入式平台的裂缝漏检率较低且满足路面裂缝检测的实时性,能够降低路面裂缝检测的成本,给出相应的维修建议,提高路面裂缝检测效率与自动化程度。

To address the problems of the current pavement crack detection methods using deep learning framework,such as hard to be applied,high cost and low efficiency,apavement crack detection system based on Jetson TX2is designed.Firstly,the pavement cracks are identified by YOLOv5network,the crack targets are segmented using U-Net network,and the pavement health is evaluated based on the segmentation results;secondly,the deep learning model is optimized using TensorRT method so as to improve the inference speed;finally,the pavement video streaming analysis system is designed and deployed to Jetson TX2embedded platform by combining with DeepStream framework.The experimental results show that the detection accuracy of the pavement crack target detection model for lateral,longitudinal and mesh cracks is more than 90%,and the reasoning speed after optimization is 30.7 ms/frame,which increases the rate by 35.1%.Finally,it is verified that the crack leakage rate of Jetson TX2embedded platform is low and meets the real-time pavement crack detection,which can reduce the cost of pavement crack detection,and give corresponding maintenance suggestions to improve the efficiency and automation of pavement crack detection.