Reserach And Implementation Of Deep Learning Based Object Detection And Fine-grained Classification For Highway

Expressway is an important part of national transportation and the lifeblood of national economic development,its smooth operation maintains the security of people's property and the development of the country's economy.With the rapid development of China's economy,the number of highways and the number of vehicles have increased sharply,which has brought great challenges to the management and operation of expressways.Although the intelligent transportation system based on machine learning has always been an important means to solve the highway management operation,but the performance of the traditional machine learning system is far from the practical requirements.With the development of deep learning technology and the implementation of the national “AI+X” strategy,the intelligent traffic system based on deep convolutional neural network shows great application prospects.Therefore,based on highway surveillance video data,this paper studies the highway object detection and fine-grained classification algorithm on the strength of deep convolutional neural network and computer vision technology,and proposes a real-time detection algorithm that accurately detects multiclass objects and calculates small consumption,meanwhile realizes an intelligent traffic analysis platform for processing multi-channel high-definition surveillance video.The main work of this paper is as follows:First,extracting data from about one hundred of high-definition monitoring sections such as the highway main line,branch line,toll station,rest area,and building a complete highway object detection dataset,including 11 types of vehicles and pedestrians,a total of55,260 pictures,520,000 object instances,the dataset covers extreme scenes such as night,rain,fog,etc.In addition,a fine-grained classification dataset of highway objects was constructed,including 11 types of vehicles and 3 types of personnel,totaling 39498 classified pictures.Second,this paper aims to research the real-time and efficient object detection algorithm based on the detection dataset,investigates and compares the performance of the general detection algorithm,then selects Faster R-CNN algorithm for optimization: Using the lightweight network MobileNetV1 as the backbone network,the second-stage feature extractor is optimized,which are used to accelerating.The paper uses eltw sum feature aggregation method,clusters the anchor boxes and adds the loss function of hard example mining to improve detection accuracy.The final model balances the detection accuracy and speed well,it can adapt to different monitoring sections,different weather and lightingconditions.Third,analyzing the characteristics of the fine-grained classification dataset and exploring the performance of the general classification network.In order to ensure the overall detection speed not be affected by fine-grained network,this paper employs MobileNetV2 and optimizes it to get 94.89% of the classification accuracy.Fourth,based on the optimized object detection algorithm and fine-grained classification model,an intelligent traffic analysis system is built to analyze the detection results,it realizes pedestrian intrusion detection function and designs object matching and tracking algorithm to analysis vehicle driving state to achieve illegal parking,retrograde,congestion,slow-moving,traffic statistics and other functions.The platform can access local video data or real-time video streaming data,and could support high-definition video real-time detection and analysis functions,it also provides practical functions such as real-time alarm and historical event query.