Reasearch On Applications Of Deep Learning Based Object Detection

In the field of computer vision,object detection is an important subject.It has significant research value and extensive application prospects in video surveillance,automatic driving,human-computer interaction and so on.In recent years,the breakthrough progress of deep learning technology and the continuous improvement of computer hardware performance had greatly promoted the rapid development of object detection technology and increasingly widely applied.The research content of this thesis is to improve the object detection method based on deep learning in some application fields.This thesis mainly involves two applications.One is pedestrian detection,and the other is isolator detection.Result from the small scale of pedestrian detection and the wide distribution of pedestrians,this thesis first proposes to fuse deep features with traditional channel features of high resolution to solve the problem that deep features are incapable of detecting small pedestrians.What's more,aiming at the multi-scale issue of pedestrians,the feature pyramid network FPN is improved.By adaptively fusing multi-scale deep features,it performs better in multi-scale pedestrian detection.For the detection and recognition of disconnectors,we have developed three deep learning disconnector detection and state recognition algorithms under different accuracy and speed requirements and put it into service.The main contributions of this thesis are summarized as follows:(1)In this thesis,a pedestrian detection algorithm based on the fusion of deep features and traditional channel features is proposed.This method extracts deep features and traditional features at the same time for every candidate regions,following with a fusion of them.With a small increase in computation,the detection accuracy is greatly improved.(2)To solve the multi-scale problem in pedestrian detection,this thesis adopts the feature pyramid model FPN.In addition,a Squeeze-and-Excitation module and an Adaptive Feature Pooling module are embedded into the network to integrate the features from different hierarchy of FPN.These two modules make the algorithm more capable of detecting multi-scale pedestrians.(3)This thesis also designs and implements a detection and identification system for disconnector.The process of data augmentation during training as well as detection and identification is customized designed.Furthermore,Different detection and recognition schemes are proposed for different accuracy and speed requirements,resulting in a favorable application effect.