Research On Object Detection And Semantic Segmentation Algorithm Based On Deep Learning

Object detection and semantic segmentation are two basic problems in the field of unmanned computer vision.It should be said that in the understanding of unmanned scenes,it is the basic ability that unmanned vehicles should have.The core task of the former is to identify and locate different categories of targets in the field of view;the latter performs pixel-level segmentation,which is more delicate,more precise,and more complex than the former.For scene understanding of unmanned vehicles and robots,research on target detection and semantic segmentation is of great significance.For these two tasks,the main content of this work can be described in the following aspects:(1)The network based on the SSD object detection algorithm and the SegNet semantic segmentation algorithm respectively re-simplifies the design.In the aspect of object detection,it focuses on several papers on the detection of RCNN series,pointing out the key points of network design in the field of target detection.First,for the acquisition of the regional candidate box,the initial sliding window to the last RPN network,to the basic convolutional network acquisition,then the feature extraction,and finally a classifier classification process.Turn the initial detection problem into a regression problem.On the network side,adding more scale integration and deeper network layer,the design of the loss function is implemented on the hardware platform of TX2.Therefore,while increasing the depth,try to adjust the parameters,compress the model,and remove the fully connected layer.(2)The training model and the segmentation model are separately trained on different data sets,and the descriptions of different data sets are expanded,the characteristics of each data set are highlighted,and the reason for selecting the training samples is selected.The training process is based on NVIDIA.The model 1080Ti GPU,as well as the Lenovo server.(3)For the simplified network model of the above design,and the training detection and segmentation model,the specific object detection and semantic segmentation algorithms are transplanted to NVIDIA TX2,and the PointGrey gray point camera is used to collect image data.The real-time image data from the PointGrey camera is used for target detection and semantic segmentation.At the same time,real-time detection and segmentation results are displayed on the display,so that the front of the vehicle can be grasped at any time,so as to improve the auxiliary behavior such as decision-making and control of the vehicle.The core content of this round is the improvement of the network.The goal detection and semantic segmentation based on deep learning have been developed to the present.Although the experimental results have been qualitatively improved,the accuracy and speed can be achieved very well.The standard,but there are still many limitations in practical applications,such as the hardware level of the computer.In deep learning,especially the deep learning training process requires high hardware configuration and high requirements for data sets.Therefore,this work tries to make some improvements based on general algorithms,based on lower computer hardware.The configuration also enables a real-time detection and segmentation effect.At the same time,it can meet the speed requirements while ensuring the accuracy,and meet the basic unmanned scene understanding needs.