Research On Flame Detection And Location Technology Based On Deep Learning

With the fast-paced advancement of machine vision and binocular vision technology,target detection and binocular ranging technology have found ample applications in scientific research,industrial inspection,and special engineering practices.Applying these technologies in firefighting robots can obtain information such as the type and distance of flame targets without contact.However,due to the variable shapes of flame targets and their small initial sizes,using object detection or binocular vision techniques alone cannot achieve accurate and timely detection tasks.Therefore,this thesis combines deep learning-based object detection techniques with binocular vision ranging techniques to propose a flame target detection and localization method for fire inspection robots.Based on this method,a flame recognition and localization system is constructed.The main research work of this thesis can be summarized as follows:(1)According to the characteristics of fire targets,the YOLOX-nano algorithm has been improved to achieve flame target recognition.Firstly,the collected fire image data is annotated using the Label Img labeling tool to create a fire dataset.Then,considering the diversity of fire targets and their small size,a lightweight YOLOX-nano object detection network is used as the base.An attention mechanism is added to improve the network’s ability to preserve small targets in the early stages.Soft-SPP is employed to replace SPP in order to retain more fire target information.Additionally,a novel loss function is proposed to accelerate network speed and improve regression performance for small targets.Subsequently,the network is trained using a pre-trained model to reduce the time spent on feature extraction.Finally,the performance of the improved YOLOX-nano model is evaluated using metrics such as average precision and frame rate,and comparative experiments are conducted.Experimental results demonstrate that the proposed algorithm can achieve real-time detection while ensuring detection accuracy.(2)Aiming at the requirements of accuracy and speed for flame ranging,a binocular vision ranging algorithm and technology were studied to achieve distance measurement of flame.Firstly,the Zhang’s calibration method and Bouguet algorithm are used to calibrate and rectify the binocular cameras as a preprocessing step.Then,a comparative analysis is performed on graph cut-based stereo matching algorithm(Graph Cut,GC),semi-global matching(SGM),and semi-global block matching(SGBM)to determine that the SGBM stereo matching algorithm is most suitable for meeting the requirements of flame distance accuracy and speed.Finally,actual binocular flame images that have been rectified are used for stereo matching experiments.The resulting disparity map in the actual scene is obtained,and the distance of the measured flame in three-dimensional coordinates is calculated.Experimental results demonstrate that the proposed method can meet the requirements of flame target localization.(3)By combining deep learning-based flame target recognition and binocular vision flame target localization methods,a flame recognition and localization system for firefighting robots has been constructed.The flame target recognition algorithm and binocular localization algorithm proposed in this thesis are implemented using the Python programming language as the core functionality of the system program.Additionally,a graphical user interface(GUI)with visualization capabilities is developed based on Py Qt5 to facilitate user control of the detection process and observation of detection results.The system is extensively tested on real flame scenes,and the experimental results demonstrate that the system can accurately and effectively identify flame targets and obtain their three-dimensional coordinate information,thus verifying the accuracy and effectiveness of the system.