Research On Visual Inspection System Of Inspection Robot Based On FPGA Acceleration

The successful operation of inspection robots relies on their ability to detect front obstacles.Visual sensors have certain advantages over other sensors in terms of cost and applicability.Moreover,with the recent advancements in visual algorithms and hardware technologies,they have provided convenience for visual detection in robotics.This dissertation presents a research study on the visual detection system for an inspection robot.The main focus is on utilizing stereo vision technology to perform depth distance detection of obstacles and using image classification to determine whether the obstacles are passable.The specific contents are as follows:Firstly,binocular vision technology is studied.A pinhole camera imaging model is established to explain the transformation relationship between the world coordinate system and the image coordinate system.The principle of ranging is analyzed,introducing how binocular vision mainly perceives the depth of the external environment through calibration and stereo matching.The Zhang’s calibration method is employed using a self-made chessboard grid for calibration of the binocular camera.Research is conducted on stereo rectification,and the transformation matrix for rectification is derived.Additionally,the stereo matching algorithm is studied,with partial improvements made to the existing algorithm and applied in subsequent algorithm implementation.Secondly,FPGA is utilized for hardware acceleration of the binocular vision detection system.A binocular vision system is built based on the ZYNQ platform,and FPGA implementation is performed for the acquisition,preprocessing,and stereo matching algorithm of the binocular images.The process is extensively researched and explained.The disparity map generated by the algorithm is stored in DDR and displayed via HDMI for validation.The system is analyzed in terms of resource usage,real-time performance,and power consumption,demonstrating excellent real-time capabilities of the FPGA-based binocular vision system.Depth detection experiments using the disparity map show that the system achieves good ranging results for objects within a certain range,effectively performing depth detection of obstacles.Next,obstacles are subjected to SVM-based image binary classification.Common image classification methods are studied,and the SVM algorithm is employed for simple binary classification of potential obstacles on the path.A SVM classification model is trained using a test set comprising common obstacles,and the model is evaluated through testing.The model is then ported to a Raspberry Pi for information exchange with the FPGA via serial communication to accomplish depth and category information detection of obstacles.Finally,experimental verification is conducted on the visual detection system.A simple walking system for the inspection robot is constructed,including the selection and control of the overall hardware system.Obstacle avoidance experiments are performed based on the visual detection system of the inspection vehicle,and the results demonstrate excellent detection effectiveness of the visual detection system for obstacles,meeting the requirements of accuracy and real-time performance.