Traffic Street Scene Recognition Based On Deep Learning

Traffic street map service as a new way,from its inception was given high expectations and attention.Traffic Street truly "human perspective",to provide users with a more realistic,richer detail map service.Street traffic is rich traffic information,location analysis and probability of various types of street traffic street in target appears is of great importance.Street traffic Street complex and diverse in background and target species varied streetscape,street blocking each other in all kinds of situations are very common goal,using traditional streetscape recognition algorithm,it is difficult to be precise target traffic classification and traffic targeting tasks.Due to mature and improve the performance of computer hardware deep learning algorithms,dealing with traffic Street traffic Street View image or video has achieved some results,so street traffic identification based on the depth of learning to become a major research direction of the image processing.This article mainly researches the deep learning street scene recognition algorithm based on candidate regions and the deep learning street scene recognition algorithm based on regression method,and studies the distribution of various targets in the actual street scene scene.The PASCAL VOC and INRIA mixed data sets are re-produced.Design and verify the operation speed and recognition efficiency of these two different algorithms.This article uses a mixed data set of the PASCAL VOC dataset and the INRIA dataset,this article uses k-means clustering method to re-determine the size and number of anchor boxes,and proves that the recognition efficiency is significantly improved through experiments.By mixing dataset Street the need to study the target data,classified according to characteristics of the street,still divided into classes and certain sports,sports and is a very important part in Street target detection.Since sports goal share data with less and less focused pixel,so that the individual target crop type,a certain percentage of the pixels for amplification,a certain proportion also increased,increased re-mixed data focus,can improve the recognition efficiency model.This article improves the original YOLO loss function algorithm.By studying the principle of the algorithm in the process of street scene target recognition,the bounding box prediction loss function of the algorithm is modified.The normalized idea of rate of change is used instead of the original loss function,so that objects of different scales are in the loss function Appears as a relative error rate.The network may effectively reduce the error due to the different sizes of the input image factors caused improved.The Dropout operation to prevent overfitting of the original network is directly deleted.A full convolutional network is used,and the 1 * 1 convolution layer is used to replace the bloated fully connected layer of the original parameter.Network improved structure has stronger generalization ability,more robust,and improved the network to speed up the convergence of neural network in training within a certain range,to improve the training speed of the model.Through traffic on the street PASCAL VOC data sets with mixed data sets INRIA dataset experiment and simulation.Due to the use of the above-mentioned series of optimizations and improvements,compared with traditional algorithms,the improved algorithm improves the algorithm’s operation speed and algorithm’s recognition efficiency within a certain range,and effectively solves the problem between algorithm’s operation speed and recognition efficiency Of balance.In this study street scene recognition,operation speed and recognition rate improved algorithm have reached optimal results.