Visual Recognition Algorithm Based On Light-weight Deep Neural Network And Its Application In Campus Unmanned Delivery Vehicle

Visual recognition is one of the core research contents in the fields of computer vision,machine learning and robotics.The intelligent operation of campus unmanned delivery vehicles is inseparable from visual recognition technology.However,due to the limited hardware equipment of campus unmanned delivery vehicles,there are certain challenges in deploying visual recognition algorithms on it.This paper takes the campus unmanned delivery vehicle as the application object,and studies the problem of driving target detection and package text recognition during its operation.By designing a lightweight deep neural network algorithm for campus unmanned delivery vehicles,an effective solution to the above two types of typical visual recognition tasks is proposed.The main work of this paper is as follows:(1)The paper proposes an improved one-stage driving target detection algorithm.For the campus unmanned delivery vehicle driving target detection algorithm,there are some problem that the network parameters is large and the algorithm is hard to train.Therefore,a dense connection structure and a cross-regional connection structure are introduced into the network.In order to solve the problem of low recognition accuracy,an attention mechanism is introduced.The improved algorithm has less network parameters.The algorithm not only ensures the real-time performance,but also maintains high recognition accuracy,which can meet the application needs of campus unmanned delivery vehicles.(2)The paper proposes a text recognition framework and realizes a three-stage neural network text recognition algorithm based on the framework.The algorithm framework is composed of text area detection,text direction detection and text recognition,which can solve the problem that the recognition accuracy of package text information is affected by the detection area and text direction.Through network selection,a fast text recognition algorithm which combines PSENet,Mobile Net and CRNN is realized.At the same time,the PSENet network is improved.The improved network reduces the network parameters and saves computing resources,even though the accuracy is slightly reduced.(3)A campus data set is established,which includes a campus driving target data set and a campus package text data set.Considering various factors such as time,weather,and road conditions,a driving target detection data set based on the campus environment is established.The data set contains 20 hours of driving video and 40,000 target annotations,which can be used for the verification of a variety of campus unmanned vehicle driving algorithms.In addition,taking into account various factors such as illumination,angle,and text direction,a package text recognition data set is established.The data set contains 5 hours of video and 1000 target annotations,which can be used to verify the package text recognition algorithm.(4)The algorithm is tested and verified in campus scenarios.Based on the campus unmanned delivery vehicle platform,the deployment of the driving target detection algorithm and the package text recognition algorithm is completed.The driving target detection and package text recognition functions of the unmanned delivery vehicle are tested in a real campus scenario,which can verify the effectiveness of the algorithms in the real scenario.