Research And Application Of The Traffic Detection Based On CNN

Intelligent transportation has received more and more attention in today’s society and has also achieved rapid development.As the core technology of intelligent transportation,vehicle pedestrian detection is widely used in various fields such as video surveillance,vehicle-assisted driving,and bionic robots.Pedestrian detection of vehicles is different from general target detection.We must have higher speed and accuracy to ensure that drivers can safely drive in a complex traffic environment.In the actual scene of real life,the traffic conditions are complicated,the weather is changeable,the target size is different,pedestrians’ attitudes and wear are different,and mutual obstruction between pedestrians and vehicles poses greater challenges for our research work.In recent years,deep-learning convolutional neural network methods have become increasingly popular in pattern recognition competitions and have attracted worldwide attention on various data sets.Most of the deep learning detection methods include three main parts: feature extraction,region suggestions,and ROI classification.The research focus of this article is to redesign the feature extraction part of the Faster R-CNN framework.The primary research substance and outstanding contributions of this article can be summarized as the below two points:(1)A multi-scale convolutional neural network model MSFF-CNN based on feature fusion is implemented.In general,the deeper the number of layers of a convolutional neural network model,the more parameters need to be trained,and the better the generalization ability of the model,but it also means longer processing time.To solve this problem,this thesis uses the pre-trained VGG16 network model,maintains the original VGG16 parameters unchanged,and adds a new network structure for fine-tuning to obtain a higher accuracy while maintaining the original processing speed.Basically all target detection systems will encounter the same problem,the detection targets with different sizes and cannot detect small targets well.Relatively speaking,the receptive fields of different size convolution kernels are also different.Large-size convolutional kernels are favorable for global features,and small ones are favorable for local features.In this thesis,combining different sized convolution kernels,the multi-scale structure is added to the original VGG16 model,which can better adapt to various targets of different sizes.For convolutional neural networks,low-level feature maps are detailed information that facilitates frame positioning.Advanced feature mapping is the semantic information,which is helpful for classification.Therefore,by combining the characteristics of the high and low layers,this thesis can obtain better detection results.(2)An improved local multiscale and global feature fusion network model GFFLMS-CNN is proposed and implemented.On the basis of existing research,according to the deficiency of the proposed model,the feature fusion structure is improved.Redesign multi-scale network structure,different multi-scales for high and low-level feature maps.And replace global multi-scale with local multi-scale for the final features fusion,it maintains the order of high and low levels of original feature maps.So we can ensure that the details and semantic information are not destroyed,and get a better fusion results.In order to evaluate the effectiveness of the algorithm,in addition to the data set KITTI specifically for pedestrians,this thesis also uses the PACAL VOC data set widely used in the target detection field to conduct experiments.The experimental results show that this thesis obtains a higher accuracy rate and maintains the speed of the original Faster R-CNN to meet the basic real-time requirements.