Research On Visual Detection And Location For Inspection Robot For Overhead High Voltage Transmission Lines

During inspecting for high voltage transmission lines,the robot should be provided with the ability to cross obstacles such as shock-proof hammer and draping clamp in order to realize large-scale autonomous walking along the lines.In order to cross these obstacles,the robot should first detect them,then identify their types and locate them.Due to such advantages as non-contact detection,abundant information perception and so on,vision sensor has become the preferred one for the robot to perceive the surrounding environment and acquire external information when navigating.It is an important task to detect,identify and locate obstacles on the ground wire by vision,so that the robot can get close to the obstacles,avoid collision and the corresponding obstacles surmounting plans are worked out according to the types of them.On the other hand,the robot should effectively detect the pose information of the wire by hand-eye vision,and feedback the information to the control system to realize the effective grasp of the wire when the robot crosses UHV towers.Based on the above requirements,the main research contents are as follows:（1）Fast detection of overhead ground wire.In order to detect the ground wire from navigation video quickly and accurately which provides the foundation for subsequent processing,aiming at the disadvantage of large account quantity of exhaustive search algorithm used in image processing,a improved hierarchical algorithm is developed.In the improved algorithm,the whole process is divided into two stages,and detection step of the first stage is determined by minimizing the time complexity.Combined with down-sampling,the detection time complexity is reduced from o（mn²）to o(m^1/2n).To test its effect,the algorithm is applied to detect the skew angle of document image and the ground wire from navigation video.In the latter detection,the normal angle range of its edge is obtained according to the expression of it.Using the priori knowledge of the angle range and that the edges on both sides of the wire are the longest two lines in the image,combined with the algorithm,the fast detection of the wire is realized.For 480×360 and 640×360 inspect images,the improved algorithm can increase the processing speed to 7.54 and 7.21 times of the traditional HT without reducing the accuracy.The average processing time is only 8.5ms and 5.0ms respectively under the condition that the normal angle detection accuracy is 0.1°.（2）Detection and identification of obstacles on the overhead ground wire.In order to detect and recognize obstacles accurately from inspection video,which are affected by camera shaking,objects motion,large-scale changes,illumination changes and so on,the dissertation combines Hu moment and HOG features of objects by PCA algorithm,and uses SVM as classifier to recognize them.Firstly,gray stretching is used to weaken the influence of illumination,and Harris corner matching method is used to detect background jitters to eliminate its influence.Then,combining frame difference image and binary image,moving objects detection is realized.After the objects were detected,Hu moments and HOG features are extracted and combined to form joint features.PCA algorithm is used to reduce them to low-dimensional features,and SVM classifier is used to classify them.Finally,images from training set and test set,including 100 and 60 images of each four kinds of fittings and one negative sample,totaling 500 and 300 images,are trained and tested respectively.When the threshold of contribution ratio is 0.85 in PCA,the recognition rate of four kinds of fittings is the highest,which is 98.3%,95.0%,96.7% and 98.3%respectively.It is better than the existing single feature recognition algorithm in literature,and the recognition time for single object is only 0.1 ms.（3）Monocular vision distance estimation for objects on the ground wire.Based on the detection of objects,a monocular distance estimation algorithm is proposed to determine the distance of it.Through the geometric relationship between the camera and the wire and the pin-hole imaging model,the relationship between the required distance and the known distance from the closest point on the wire to the camera lens in the image is obtained.In order to test the effect of the algorithm,the experiment for the camera in both static and motion conditions is carried out.In the static experiments,the absolute and relative error do not exceed 0.20m and 10% respectively in the distance range of 1.5-9.5m and the mean absolute value of the latter does not exceed3%.In the test for the camera in motion condition,the sums of the estimated distances and these traveled by the robot are basically a fixed value,which meets the expectation.Through the two experiments,it is proved that the method is reliable.The method has such advantages as higher detection precision,less parameters needed and so on.（4）Automatic calibration of the navigation camera.Aiming at the difficulty of camera calibration on overhead ground wire,on the basis of Zhang’s calibration algorithm,an automatic calibration method is designed,which can complete the calibration without taking images of the calibration board manually and obtain parameters such as focal length.In calibration,the robot maintains a low uniform motion at a known speed after detecting a moving object on the wire.In the process,the camera pitch angle remains unchanged.Through the object in n images,3×n points of intersection are obtained from the wire,and a checkerboard image with 2×（n-1）rectangular of the same known size is constructed.These points are generated by the intersection of both edges and axis of the wire and their perpendicular line located in the vertical plane of the axis.Three chessboard images with the same number of rectangles are acquired by changing two deflection angles,and the calculation was completed in the same way as Zhang’s calibration method.When the average focal length is 1420.27,2524.69 and 1855.23 pixel respectively,the average error of automatic calibration is 4.41%,5.72% and 8.31% based on Zhang’s calibration results.（5）Detection of the pose parameters of the ground wire in hand-eye visual images.In order to accurately detect the slope angle and intercept of the wire from hand-eye vision images,a method is designed to determine objects segmentation area by clustering the straight lines according to the characteristics of the wire in the image.Through the edge image,the normal angles,the distances from the origin and starting and ending points of the lines are obtained and these lines which are nearly parallel and close to each other are regarded as the same category by clustering.After removing the outer lines,the segmentation region of each object is determined by the endpoints of the lines in the same category,the slope angle and intercept of the center line of it are obtained,and the region is separated as an objects to be recognized.Finally,the LBP features of the objects are extracted,and SVM classifier is used to recognize the ground wire from the objects.Experiments show that the maximum error of slope angle and intercept are 3.4°and 9.5 pixels respectively,and the latter is less than 1/24 of the image width.