Study On Performance Improvement Of Hidden Danger Detection System For Transmission Line Based On Image Sequence

As power enters into all walks of life in society,accidental power failure caused by hidden danger such as foreign object invasion and fire seriously affects people's lives,and which also results in enormous economic loss.Therefore,for normal power supply,how to effectively monitor the high voltage transmission line and its surrounding environment,so as to timely detect hidden danger,has become an urgent problem to be solved.According to statistical data,common hidden dangers around transmission line include engineering vehicle,telescopic boom lift,tower crane,fire,foreign object,etc.At present,aiming at hidden danger detection,we mainly classify the existed methods into four categories: manual inspection,patrol robot,unmanned aerial vehicle inspection and video surveillance based on camera installed on tower.Although the above methods have achieved certain effects,the first three methods not only require manual intervention,but also have limited monitoring scope,and it is difficult to realize real-time surveillance.For video capturing,it requires a better performance for battery.However,transmission line cannot supply power to the camera directly,and rechargeable battery is used in general cases.Rechargeable battery usually utilizes the wind energy and solar energy,and monitoring is realized by capturing the image sequence at interval,which can further decrease the power consumption.In order to detect hidden danger via image sequence,research group began to cooperate with a company in Shandong from 2015,and we combine the theory of image processing and machine learning to analyze the image sequence.For the camera,the monitored scene is fixed.Thus,image difference is used as a major detection technology,so as to determine the location and type of hidden danger,and alarm according to detection result.At the end of 2016,the first version of hidden danger detection and recognition system is developed.Although the system achieves the detection of the hoist boom in sky region,the engineering vehicle and fire detection in ground region,it still has some problems to be solved: it cannot detect the foreign object invasion and tower collapse in sky region;there is still some noise in the result of image difference;the number of hidden danger type that can be recognized is very limited,and it cannot judge the hidden danger such as tower crane,pump trucks,etc.Aiming at the problems in first version of system,this paper mainly carries out the following works:(1)Segment skyline.The accurate position of skyline is obtained according to the location of sky and the degree and frequency of gray-level change.Then,the mask of sky region is obtained.Because the types of hidden danger in sky and ground are different,the sky mask is applied to segment image into sky and ground,and hidden danger is detected separately.(2)Propose the cumulative edge probability distribution model,and realize the detection of hidden danger in sky region.Hidden dangers in sky region include tower collapse,foreign object invasion and camera abnormal movement.In fact,the probability of occurring the above hidden dangers is small.Because the tower and transmission line are fixed targets in sky,although their positions change occasionally,in general cases,they just shake near their equilibrium position.Since the monitored scene is fixed,from the perspective of image sequence,these fixed targets should conform to a certain probability distribution model.Based on a large number of observations,a cumulative edge probability distribution model is proposed,and it is combined with adaptive Parzen window.Then,hidden danger is detected by comparing the model.(3)Detect the hidden danger in ground region.Firstly,the improved multi-color space fusion algorithm is applied to obtain changed regions;secondly,these regions are filtered by combining the gray-scale fluctuation shift,improved multi-scale edge orientation histogram and fractal dimension;finally,the type of hidden danger is determined by the features of texture and shape;(4)Recognize the crane and engineering truck.In the view of complementarity between the Faster-RCNN and traditional image processing methods in detecting hidden dangers in ground region,the false negative rate is reduced.At present,the system has been promoted in Jiangsu,Liaoning and 17 cities of Shandong.The system developed by our group has reduced 98% workload of the staff according to statistics.Although the algorithms in this paper have further improved system performance,and achieve better detection result,it is not satisfactory in object occlusion and small targets.Thus,the system performance need to be improved in the future.