Research On Traffic Density Estimation And Vehicles Detection By Using Convolutional Neural Networks

Intelligent Transportation System is one of the core components of Smart City,and can do significant contributions for alleviating the traffic congestion and help improving traffic efficiency.With improvement of technology and hardware,traffic analysis is relying on data collected from various traffic sensors nowadays.Among various traffic surveillance techniques,computer vision based video surveillance systems are often used for monitor-ing and characterizing traffic load.Knowing in real time when the traffic is fluid or when it is jam is a key information to help authorities re-route vehicles and reduce congestion.The detection of vehicles pictured by traffic cameras is often the very first step of video surveillance systems,such as vehicle counting,tracking and retrieval.Most traffic scene analysing methods are based on motion features,such as background subtraction,optical flow,dynamic texture features.Besides,lots of vehicle detection algorithms are still using hand-craft features,and can only deal with surveillance data from a specific check point.There are lots of challenges in traffic surveillance videos analyzing,such as limited open-access data,low frame-rate and low-resolution videos,various shooting angles among different surveillance cameras,challenging weather and lighting conditions.In this thesis,we use the deep-learning based methods to deal with these challenge situations in the intelligent traffic analysing,and the work in this thesis can be divided into two parts:Traffic density analysis in traffic scenes and Vehicle detection and classification.The main content and contributions of thtis thesis can be summarized as follows:1.Since we can not compute the motion features for long-distance filmed low framer-ate videos,as well as cannot detect vehicles due to small scale and occlusion issues.In this thesis,we proposed to use the Convolutional Neural Networks(CNN)to analysis the traffic density of a traffic scene from whole view.Firstly,traffic density is defined as the percentage of road being occupied by vehicles in an image.In our previous work[1],we validated that the traffic status is highly correlated to its texture features and that CNN has the superiority of extracting discriminative texture features.We proposed several dif-ferent CNN models to segment traffic images into three different classes(road,car and background),classify traffic images into different categories(empty,fluid,heavy,jam)and predict traffic density without using any motion features.To generalize the model trained on a specific dataset to analyze new traffic scenes,we also proposed a novel trans-fer learning framework to do model adaptation.2.The large-scale ImageNet dataset has made tremendous contribution for the success of deep learning in recently years.Realizing the importance of large dataset for intelligent traffic analysis,we built and released the largest traffic surveillance dataset(MIOvision Traffic Camera Dataset(MIO-TCD))in the world for vehicle localization and classifi-cation in collaboration with colleagues from Miovision inc.(Waterloo,On).With this dataset,we built an online evaluation system and organized the Traffic Surveillance Work-shop and Challenge in conjunction with CVPR 2017.We evaluated several state-of-the-art deep learning methods for the classification and localization task on the MIO-TCD dataset.Considering the results,we may conclude that state-of-the-art deep learning methods ex-hibit a capacity to localize and recognize vehicle from single video frames.While with a deep analysis of the results,we also identify scenarios for which state-of-the-art methods are still failing and propose concrete ideas for future work.3.Saliency detection aims to highlight the most relevant objects in an image(e.g.vehi-cles in traffic scenes),we proposed a multi-resolution 4*5 grid CNN model for the salient object detection which is inspired by the Mumford-Shah model[2].The model can effi-ciently fuse the local and global feature for saliency detection,and is trained based on a novel loss function.We tested the model on several saliency benchmark datesets,and re-sults showed that our model enables near real-time high-performance saliency detection.We also extent this model to do traffic analysis(e.g.foreground vehicle segmentation),ex-perimental results on the MIO-TCD dataset demonstrated that our model can do precisely foreground vehicle segmentation.At the same time,we further tested the model trained on MIO-TCD dateset on various traffic videos,whose results showed that the model has a strong generalisation ability.In conclusion,it is possible to analyze the traffic density status by the CNNs without using motion features.Experiments on large scale traffic dataset demonstrated that well-trained deep models can do precisely vehicle detection and classification,and these models can generalize well for analyzing various traffic scenes.