Research And Application Of Target Detection Method Based On Learning

Target detection is the basis of target tracking,recognition and so on.The quality of its method is closely related to the accuracy of tracking, recognition and processing.In real life,due to the influence by different illumination,local occlusion,target scale and rotation changes and other factors during the process of image or video acquisition,results in the appearance of the target morphological characteristics has changed a lot,which has brought great challenges to the target detection.Aiming at the main problems of target detection in both static scene and dynamic backgrounds,this paper mainly studies the target detection method based on learning.Extracting power lines in high voltage transmission lines and monitoring the transmission line,the deformation and the trouble of electric power towers are main tasks in automatic power inspection.In the helicopter power inspection,detecting towers from videos plays a vital role in the determination of tower type and the judgment of tower deformation and trouble.This paper presents a two-stage tower detection method based on learning,which consists of following steps:First,cutting rich tower and non-tower pictures in transmission line inspection videos taken by helicopter to construct the training sample set and to mark it.Secondly,the Local Binary Pattern(LBP) feature is extracted from training sets.Send feature sets into the Adaptive Boosting(ADABOOST) model for training and generating the classifier1.Then,the depth of learning Convolution Neural Network(CNN) model structure is designed.Send training sets and annotation information into the CNN model under the Convolution Architecture for Fast Feature Embedding(CAFFE).And the classifier2 is generated.Finally,in the multi-scale,the test video image blocks in the sliding window are sent into classifier1.And the candidate regions of the tower are obtained according to classifier1.Sent regions into classifier2 to judge whether it is tower.According to the classifier2's output to get the accurate positioning of towers.And the accurate positioning of the tower is obtained from classifier2.The service robot for acquiring text information in the natural scene has a wide application prospect in the field of blind auxiliary navigation and visual positioning.Because of the diversity of characters in location,font,color and the ambiguity,occlusion and other factors in natural scenes,the detection and positioning of texts in natural scenes becomes a very challenging problem.This paper presents an efficient method for text signs detection based on Bag of Visual Word(BOVW) model,which consists of two parts:training and testing.In the training part,firstly,the Binary Robust Invariant Scalable Keypoints(BRISK),which is simple to calculate and has a certain robustness to scale changes and rotation,is chosen as the texture feature of text signs.Secondly,the BRISK feature of images is extracted and Self-Growing and Self-Organized Neural Gas network(SGONG) clustering is used to obtain the visual dictionary.Thirdly,the BRISK feature extracted from positive and negative images is quantified by the dictionary to obtain the BRISK shape histogram and the HS color histogram is extracted and calculated further to obtain the CIHS histogram.Fourthly,Combining the two features,then the strong differentiation feature of text signs is obtained.Finally,the text signs sample set is trained by ADABOOST algorithm to get the text signs detector.In the test section,firstly,the Maximally Stable Color Regions(MSCR) algorithm is used to detect the candidate regions of the text signs in the natural scene to reduce the complexity of using the classifier to detect directly.Then,the text signs detector obtained by learning is used to finely detect the text signs in the candidate regions to get the positioning of the text signs.The detection method of the electric power tower in the helicopter power inspection system in this paper is tested with test videos that is totally about 30 minutes.The results show that the precision,the recall and the F1-measure of the method can reach 92%,79%,and 85%.Its detection time of average per frame is about 0.33S.Compared to the cascade ADABOOST method and the deep learning CNN method,the tower detection method have better detection performance.Thus,the tower detection method can be applied directly to the automatic detection of the tower in the helicopter power inspection system in order to further perform the maintenance task and make fault diagnosis.The text signs detection method in the natural scene in this paper is tested with 678 street view images(including 661 text signs).The results show that the detection rate of the method reaches 76%,81%,90% respectively,and the sensitivity of the method is 58%,78% 89% respectively for the far, medium and near text signs.Compared to approaches using CIHS Histograms,Scale Invariant Feature Transform(SIFT), SIFT+HS,BRISK,Fast Retina Keypoint(FREAK) or FREAK+HS feature,the text signs detection method have better detection performance for its more accurate positioning,less false detection,higher detection precision and less detection time.Therefore,the text signs detection method can be applied directly to the detection and location of the text signs in the natural scene in order to further perform the text segmentation and recognition.