Research On Target Detection And Location Method For Vision System Of Intelligent Mechanical Arm

Due to its high efficiency and flexibility,the robot has been widely used in industry.In recent years,with the development and research of machine vision,robot vision system has become one of the most promising fields of application and theory research because of its large amount of information and wide application.In order to adapt to the actual engineering application,enhance the manipulator operation ability of the complex work environment,and enhance the level of the intelligence and autonomy of the mechanical arm,robot visual system needs to be developed,so as to achieve the perfect visual servo function.In this paper,according to the practical engineering application,the target detection and location method of vision system is studied and explored.The main work includes the following aspects:1.According to the research,the target detection and localization methods are summarized,and the basic principles and key technologies of the mechanical arm vision system are introduced briefly.2.A multi-object detection method based on single template is proposed.The traditional target detection algorithm can only identify a single target,and the multi-target detection method requires a long time to learn the template or feature.Most of the algorithms can not guarantee the efficiency with high accuracy.In this paper,a method based on NCC algorithm and NMS method is proposed,which is precise,afficient and robust to noise and illumination changes.3.A multi-object detection method by rotation-invariant HOG feature based on FFT is proposed.Traditional template matching method is only applicable to the scene that the target and the template image have a slight angle rotation or no rotation.In the presence of rotation angle of the scene need to use the characteristics to detect the target.Based on the rotation invariant HOG features based on polar Fourier analysis framework proposed by Liu,this paper raises a new 108 dimensional rotation-invariant HOG feature in Fourier space.And then it is accelerated based on the FFT method,combined with the SVM classifier to detect the target,and get a good detection effect.4.A multi-object detection method based on edge information is proposed.In practical engineering projects,the target detection method based on image matching and single classifier is not suitable for the detection of different targets.In this paper,the Edge Boxes algorithm proposed by Microsoft Research Institute is improved with adding the shape constraint and the redundant target position is eliminated by using the NMS method.Setting the parameters involved in the algorithm reasonablely can have good target detection results.5.In the paper,a simulation experiment is carried out by a simulated mechanical arm composed of fixed two cameras.The experiment consists of two parts: hand-eye calibration experiment and pose measurement experiment.The hand-eye calibration experiment is carried out by using TSAI's method for the calibration of the hand-eye system of mechanical arm.The pose measurement experiment is carried out by using PNP algorithm and binocular intersection to measure the position and pose of the object in the scene.The experimental results show that TSAI's hand-eye calibration method and PNP algorithm and binocular intersection are suitable for the actual task of picking up the target object by the mechanical arm with high precision.They can be widely used in the vision system of the mechanical arm.