The Research And Development Of The Object Detection And Subdivision System Based On TensorFlow Framework

The rapid development of computer science makes human life more and more intelligent.Artificial intelligence has always been an important area.The computer vision has always been an important research direction of artificial intelligence,including image processing,machine learning,pattern recognition and other disciplines.The target detection is an important research area in computer vision.Target detection and subdivision is mainly to detect the target in the picture.For example,the target detection requirements of the traffic video in the Intelligent Transportation System is to detect vehicles and the target detection requirements in the situation of manless driving is to detect pedestrain,vehicle,cyclist,traffic light.It has higher requirements on the detection and subdivision system as for the complexity on real scene.Most of the traditional object detection algorithms extract calssic artificially-designed feature and classify them with classic classifier.As the use of manual characteristics,so poor robust,and the workload is large.Meanwhile selected area strategy based on the sliding window is redundant and complicated.Instead of the traditional algorithms based on convolution filtering,combined with the deep learning,we propose an object detection method to detect objection and subdivide in the road.This paper is based on TensorFlow deep learning framework,which is an open source software library for machine intelligence and an learning software library for realizing neural network framework.This paper adopts the classic object detection scheme based on deep learning: R-FCN and SSD.This paper adopts the semi-automatic annotation method,combined with the incremental learning,to label large amounts of data.With using the calibration data to train the model,set the parameters,a model can be trained and applied to the detection system.The object detection and subdivision system in this paper contains three modules: data annotation module,model training module and detection module,tested and analyzed the model over various algorithms.As the test results shown,the object detection algorithm based on deep learning could detect and subdivide effectively the various kinds of vehicle including hatchback,sedan,truck,bus and van in the situation of traffic scene and pedestrain,vehicle,cyclist,traffic light in the situation of manless driving and the SSD could detect videos in real time.