Traffic Object Perception Technology And Its Application Based On Deep Learning

In the past few decades,artificial intelligence has made huge progress in various industries.Autonomous driving is an important branch of artificial intelligence and has attracted widespread attention.Autonomous driving technology involves a wide range of fields.As the eye of autonomous vehicles,the environmental perception system needs not only to timely process the information returned by various sensors,but also to transmit the results to the decision system to assist the vehicle to make correct judgment.At present,artificial intelligence modules such as parking assistance,emergency braking and abnormal detection can no longer meet the expectations on current vehicles.Many famous companies are striving to build fully autonomous vehicles,but other companies are also studying more economical artificial intelligence modules to assist driving.How to configure these modules in the corresponding hardware devices and integrate into the existing vehicle assistance system without losing performance is a thorny problem.The actual traffic conditions are complex and changeable.For autonomous vehicles,how to efficiently and accurately understand the external environment of the vehicle is a great challenge to the environment perception system.Therefore,in order to enhance the safety of environment perception system,this thesis studies the road traffic object detection algorithm,semantic segmentation algorithm and object tracking network model.The main research contents are as follows:(1)The real-time traffic object detection algorithm under complex traffic conditions based on deep learning is studied.The advantages and disadvantages of the current lightweight backbone network are explored and compared.The difference between the ordinary convolutional layer in the backbone network and the lightweight network Ghost Net is proved.The attention mechanism is embedded to fuse more spatial features,and the importance of each feature channel is evaluated.The advantages of the DW convolution compared with the ordinary convolution are verified.While reducing network parameters,the real-time performance of the network is improved.The BDD100 K datasets is improved to achieve accurate detection of all common traffic objects in a total of 13 categories including traffic signs and lights.Experimental results show that the network yields a model file size of 5.22 MB while achieving 15.58 FPS on real-time performance on Inter(R)Core i5 CPU.It has the ability to achieve a balance between detection speed and accuracy,and is easy to deploy to the terminal.(2)The real-time traffic object semantic segmentation algorithm based on deep learning is studied.The structural characteristics of the current semantic segmentation models are explored.The algorithm framework based on encoder and decoder structure is determined.Aiming at the problem that the semantic segmentation network is difficult to achieve a balance between speed and accuracy,a residual aggregation module is designed in the encoder to simplify the model and effectively extract features.In the decoder structure,transpose convolution and ordinary convolution are combined to gradually restore the image pixel information.The attention mechanism pyramid pooling module is designed to enhance the feature information flow between encoder and decoder,which greatly improves the network performance.The joint training on the Cityscapes and BDD100 K datasets is proposed to improve the model robustness and make the model have better performance.Experimental results show that the proposed EDNet network model obtains 47.2% m Io U on the Cityscapes and BDD100 K datasets,FPS reaches 23,which can achieve a balance between speed and accuracy.(3)The real-time traffic object tracking algorithm based on deep learning is studied.The object tracking algorithm flow is analyzed.The detection based tracking framework is constructed.In the process of multi-object tracking,it is difficult to track the object accurately due to the change of object scale and posture,and the identity occlusion.The m-yolov5 s detection algorithm is proposed as the detector of the object tracking network.The k-means algorithm is used to cluster the label into different sizes anchor.The threescale prediction is improved to four-scale prediction,which ensures that accurate anchors are allocated on different sizes feature maps.The Hungarian algorithm is used to match the trajectory predicted by Kalman filtering with the detection object.The similarity matrix is constructed to find the corresponding object between frames and track multiple object in real-time.The experimental results on the MOT16 benchmark show that the whole algorithm has strong robustness and can achieve multiple object tracking under complex traffic environments.