Study On The Method Of Target Recognition In Complex Environment

Target recognition technology is a technology to separate and identify a target from other targets in a specific environment.Although researchers from various countries have been studying this technology for a long time,it is still a main subject of the research.The target recognition technology is also a basic technology of the robot.The complex environment has put forward new requirements for the target recognition performance of the robot.Aiming at studying a target recognition method in complex environment,Deep Learning method is combined with image processing method in this thesis and a human target recognition system based on the Convolutional Neural Network is built.And the performance of the optimized model is tested on this system.The contribution of the thesis can be summarized as follows:(1)A scheme of human target recognition based on the Deep Learning model is proposed.Through the investigation and comparison of the related technologies of target recognition,the Caffe framework and the classic target detection model Fast-RCNN are taken as the main method.Then the image segmentation and positioning technology are usesd to improve the existing models.Using this model,a target recognition system is constructed to achieve the recognition of the human targets.(2)The existing VGG network model is improved and a new Convolution Neural Network model is built.First,the image segmentation technology is used to improve the Region Proposal Network(RPN)and the improved network is named as NRPN.Second,the NRPN is added to the VGG network as an improvement of the existing model.The new model will realize the segmentation of the image,the feature extraction and the classification of the targets.(3)Using the improved model,a simple human target recognition system is established to test the performance of the new model with the help of the open source tool Caffe framework.In the model training and testing process,a new data set consisting of 11900 pictures is built.To make the experiment more reasonable and satisfactory,the proportion of positive and negative samples is decided with full account.Most of the pictures in the negative sample database are selected from the pictures taken randomly in life.The results confirm that the method proposed in this thesis realise an optimization in recognition rate,missing rate and speed to some degree.And most human targets in complex environment can be recognised by using the method.