Research On Object Detection Technology Based On Convolutional Neural Network

Computer vision is an important way for the machine to imitate the human eye to understand and obtain the rich information in the image to perceive the external world.Among them,recognition and semantic segmentation are the two most important problems in computer vision.Due to the development of deep learning,especially convolutional neural network(CNN)in recent years,brings convenience to solve these two problems.Existing CNN methods have achieved superior performance in multiple tasks such as image classification,object detection and semantic segmentation.However,there is still room for improvement in these CNNbased methods in engineering applications for some specific occasions.For example,for some specific application scenarios,the general CNN method is not effective.Compared with traditional methods,it consumes more computing resources,so it is difficult to deploy on embedded platforms.and Poor performance under bad external environmental conditions.Therefore,this paper has done the following main works for the two specific engineering applications of target recognition in helmet wearing detection and semantic segmentation in lane line detection:First,in the helmet wearing detection scene,in order to solve the problem that the open source helmet data set on the Internet is very few and the quality is poor,this article collects network pictures and collection pictures from real scene,and adds annotations to construct a novel helmet wearing detection data set.In order to solve the problem that it is difficult to achieve high-accuracy detection and real-time detection at the same time in the detection of helmet wearing in the current building construction scene,this paper implements an efficient helmet wearing detector by improving the YOLOv4 network and using the NCNN framework for reasoning acceleration.The experimental results show that the detection accuracy of the detector is higher than the existing methods,and the application in actual construction scenarios also has superior test results.Secondly,in the lane line detection scene,because the traditional CNN method is not suitable for the narrow and long lane line curve detection scene,this paper proposes a lane line extraction framework based on a semantic segmentation two-branch network.Aiming at the problem of large errors in the traditional lane line fitting method when the image is converted into a bird's-eye view,this paper trains a special convolutional network to fit the lane line to solve this problem.The experimental results show that the lane line detection system constructed in this paper can obtain good detection results on the tusimple data set.