Research On Orientation Detection Of Objects On RRU Model Based On Deep Learning

The position and orientation of objects is hard to detect.Many classical algorithms are not accurate enough.The deep learning methods can achieve much more excellent performance than classical ones.Some deep learning models construct two-stage network,one to detect position and the other to estimate orientation by clip out the sub-image.But they slow down the speed of detection.Another models use one-stage deep neural network to detect position and orientation directly.This paper mainly focus on acquiring the position and orientation of power port and fiber optical port of Remote Radio Unit(RRU).The orientation of objects are detected based on position of four directional key points which are four corners of objects.Then an one-stage neural network is proposed to detect position of four key points called Single Shot Directional Points Detector(SSPD).In order to be beneficial to predict position,I use eight parameters to encode bounding boxes and propose rotated bounding boxes structure which can improve precision of detection.After that,for calculation of matching between two bounding boxes,I use a new matching measurement.This paper proposes a method to transform the position of key points into orientation degree.Following network design,this thesis narrates the data sampling,data augmentation and data labeling.Then I propose a loss function to control the network learning gradient and use SGD to train my model.In experiment,the performance of SSPD is excellent,and it can achieve 71% m AP on my own dataset consist of 50000 samples.Because the network is applied based on deep learning structure Caffe,this paper introduce how to transform the idea into codes.Finally,I do some comparative experiments.In the first part,I use different training choices and analysis the detection results.Then I apply classical methods which are Hough Transformation and Minimum Bounding Rectangle to detect orientation of objects.Compared to classical methods,deep learning can overcome the disability easily and achieve higher detection precision.