Multi-Object Detection And Classification Based On Deep Learning In Traffic Scene

The rise of artificial intelligence provides a new solution for traditional transportation problems.As an important branch of artificial intelligence,the application of deep learning in transportation has become a research hotspot.This paper focus on the multi-object detection and classification problem in traffic scene by using deep learning.By analyzing the current object detect model and classification model,this paper constructs a traffic object detection and classification network,and implements the traffic scene multi-object detection system and target classification system.The main content of this paper are as follows:Firstly,the basic models of deep learning are introduced,including deep neural networks,convolutional neural networks,deep belief networks,recurrent neural networks,and generative adversarial nets.In this paper,the convolutional neural network is mainly used,so it is analyzed,including the composition of convolutional neural network,the characteristics and training mechanism of convolutional neural network.Secondly,for the traditional target detection algorithm,there are many problems such as complicated procedures and poor environmental adaptability.This paper designs a multi-object detection system in the traffic scene which is based on deep learning.The system uses the RetinaNet object detection model and for needs of the traffic scene this paper makes some improvements as follows:it deepens the convolutional network from 101 layers to 152 layers,resulting in increase of mean Average Precison(mAP)of traffic scene objects(including cars,pedestrians,traffic signs,traffic lights)by around 2.1%;the model was accelerated and compressed,while the average detection rate was almost not lost(compared with the RetinaNet(ResNet50)model),the training time of step was reduced by 21%,and the model size was reduced more than one time.Thirdly,based on the object detection system,this paper designs a traffic object classification system.The system includes a traffic sign classification system and a vehicle classification system.The SignNet based on convolutional neural network was designed to classify the traffic sign.Compared with the previous best network,the accuracy increased by 1.08%,reaching 98.2%.The vehicle sub-classification system uses the Billnear CNN model.Based on this,the data enhancement technology is used,the system could classify 210 vehicles.Compared with the VGG-19,the classification accuracy is improved by 5.2%.It reached 90.1%.Finally,this article tests the traffic scene multi-object detection system in real video.The test result shows that the speed of detection is 7fps.The detection accuracy of the target vehicle,pedestrian and traffic light are 81.28%,42.38% and 39.63%.