Road Scene Vehicle Detection Method Based On Deep Learning

With vehicle and road as main carrier of transportation,mastering its wide-range and full-road area as well as distribution features can help to obtain traffic information in a real time way and optimize intelligent transportation system,being of very important significance for urban planning and construction of smart city.In consideration of high monitoring cost for current ground sensor,limited monitoring scope of traffic information and other deficiencies,traditional detection technology has failed to satisfy demand of wide-range traffic information monitoring.In recent years,rapidly developing modern remote sensing technology has brought new thought and new means to traffic information monitoring.Emerging of aerial remote sensing technology makes it possible for people to obtain a lot of high-definition aerial image data and realize identification and positioning of wide-range traffic information.To efficiently and accurately judge vehicle objective from image,deep learning algorithm is combined with remote sensing technology in the Paper and a Faster R-CNN target detection method based on road area is proposed regarding UAV image vehicle detection.Firstly,UAV image is provided with road area extraction.The extraction of road area-of-interest can reduce false alarm not on road during vehicle identification and greatly reduce calculated amount of data to be detected.On this basis,Faster R-CNN convolutional neural network model is used to realize detection and positioning of vehicle target.Main work is as follows:(1)First off all,several main deep neural network algorithm are introduced.The principle and structure characteristics of deep convolution neural network is demonstrated and analysed,which shows that the deep learning method has unique advantages in digital image recognition,as well as a summary of the relevant work of the target recognition framework in vehicle detection in recent years.(2)Extraction of potential road area for UAV image vehicle.Firstly,deep convolutional neural network method for semantic segmentation is proposed on the basis of deep learning theory and through the construction of a fully convolutional network classification model to achieve UAV image road extraction.The network training is implemented by the Python language docking Mxnet framework.Through continuous iterative learning,feature expressions at different levels of the sample data are obtained.The entire training process is iterated for a total of 70000 times.Finally,the resulting network model can achieve high-precision road extraction.Compared with object-oriented classification method based on manual extraction of various sample features,the training process is easier.Through experimental comparison,road extraction precision of convolutional neural network model is improved by 10%than traditional object-oriented method classification precision and Kappa coefficient increases by 0.11.In addition,road extraction method of convolutional neural network reflects end-to-end and data driving thought without need of special design features and reducing manual participation steps to the greatest extent.It is drastically improved than object-oriented method in generalization performance and automation level.(3)A vehicle detection method for Faster R-CNN deep convolution network based on road area constraint is researched and single-target&multi-target vehicle detection models are built.Faster R-CNN vehicle detection network model is subject to sample training mode for learning network model parameters.The whole network training process is realized through linking python language with Mxnet framework.With original spectral value as input,transfer learning mechanism is introduced and a few samples are used for slight adjustment and training to effectively build deep convolution network model oriented at UAV data vehicle identification.380 images were tested in the experiment and the whole time was less than 3 minutes.Through verification,mAP values for test results of Faster R-CNN single-type and multi-type vehicle target detection model respectively reach 0.904 and 0.843,effectively solving UAV image vehicle detection problems under complex scenarios.