Design Of Embedded Concrete Crack Real-time Detection System

Crack is a common defect in concrete structure,which seriously threatens the safety and stability of the building.Regular crack detection is essential to ensure structural safety,prolong service life and promote engineering development.However,the traditional manual detection method is time-consuming,inefficient and inaccurate.The detection method based on digital image processing technology relies on manually designed feature extraction rules and lacks flexibility and self-adaptability.With the gradual enhancement of intelligent analysis and parallel processing capabilities of terminal AI chips,deep learning deployment on embedded platforms and application in the field of detection has become a hot topic.At present,foreign AI products still dominate the market.In order to reduce the dependence on foreign technologies and promote the development of local chips,it is of great significance to develop embedded real-time crack detection system based on deep learning for domestic intelligent platform.In recent years,deep learning algorithms have developed rapidly.These algorithms have the advantages of automatic learning,extraction and fusion of image features,so they are gradually applied in the field of crack detection.However,the deployment of these algorithms requires high computing power of the platform,and it is difficult to realize real-time deployment of traditional embedded devices.With the continuous enhancement of intelligent analysis and parallel processing capabilities of terminal AI chips,it is a new prospect to deploy deep learning algorithms on artificial intelligence platforms and apply them to crack detection.At present,foreign AI products are dominating the market.In order to reduce the dependence on foreign technology and promote the development of local chips,it is of great significance to design a crack real-time detection system based on domestic intelligent platform for deep learning algorithm deployment.The main contents of this paper are as follows:1.This paper compares representative domestic terminal AI chips and selects RK3399 Pro as the core of the intelligent platform to build a stable,efficient and referential deep learning development and deployment environment.2.Designed crack detection algorithm for artificial intelligence chips with limited computing resources.The overall framework of the algorithm is divided into two stages: crack recognition and semantic segmentation.In the crack identification stage,several lightweight networks are fine-tuned,and the generalization of the model is improved by combining data enhancement technology.According to the experimental results,the identification model suitable for embedded platform deployment is selected.In the semantic crack segmentation stage,the lightweight improvement and optimization of U-Net network,combined with the training strategy of combined loss function,get a relatively high precision lightweight segmentation model.3.With RK3399 Pro as the core,an embedded real-time crack detection system is designed which can be mounted on the UAV to complete the inspection task.Image acquisition,image preprocessing,crack detection and image post-processing modules are planned in the system.In terms of hardware design,the camera selection and circuit design of several modules have been completed,and the size and weight of the equipment meet the requirements.In the aspect of software development,combined with the characteristics of the platform chip,based on QT cross-compilation environment for the corresponding programming and software optimization.Finally,the performance of the system is verified in the three aspects of algorithm model deployment,modularization and integrity.The test results show that the system has good overall stability,the overall processing capacity of 35 FPS,can meet the real-time detection requirements,and can better identify and partition cracks.