Research On Vehicle Detection And Attribute Analysis Method Based On Deep Convolutional Neural Networks

According to the latest authoritative statistics,the number of motor vehicles has reached 304 million,and the number of drivers has reached 371 million in China.The national highway has a total of 131 thousand kilometers,ranking the first around the world.China has entered the "Automotive age" in big strides.Facing the rapid development of the huge traffic system,people's demand for improving the performance of Intelligent Transportation System(ITS)solution is becoming increasingly urgent.ITS is composed of a series of wired or wireless technologies including information,electronic and control.While these technologies are integrated into the transportation infrastructure and vehicle,they can help to monitor and manage the traffic flow,reduce congestion,provide optional route,ensure the safety and save time cost for travelers.For ITS,the technologies about vehicle detection and vehicle identification directly influence the feasibility and the applicability of it.Since vehicle images collected from the real-world are easily affected by illumination,occlusion,viewpoint,distortion and so on,it is more difficult to improve the accuracy rates of vehicle detection and vehicle recognition.This dissertation has carried on the thorough research to the above several problems.The vehicle detection method under the invariant scene is studied in this dissertation.Aiming at the problems of vehicle loss and template updating not timely,which are caused by the fuzzy brightness between targets and background,an adaptive template library with real-time updating is proposed.Firstly,the continuous sequence images extracted from video are processed with differential operation to get several targets.Secondly,these targets are compared with templates from library by calculating the maximum normalized correlation between them.Finally,the target with the highest similarity is extracted as vehicle.Corrosion calculation,sparse processing and full matrix operation are used to improve the real-time performance.The experimental results show that the proposed solution can effectively detect the vehicle in the complex environment in real-time.Moreover,in order to solve the limitation of invariant scene in the above,the method based on the feature extraction is applied in this dissertation,to research on the vehicle detection method based on Deep Convolutional Neural Networks(D-CNN).Firstly,in the image pre-processing stage,the pixels are merged into blocks according to certain rules,and then these blocks are merged into boxes according to the similarity between them.The merged box is a rectangular box that may contain a vehicle in this specified area.Secondly,under the Caffe(Convolutional Architecture for Fast Feature Embedding)framework,D-CNN is used to extract the feature of each merged box.Finally,the LibSVM classifier is used for final judgment to complete the vehicle detection.In this dissertation,the proposed method is compared with the classical Histogram of Oriented Gradient(HOG)method and so on.The experimental results show that the proposed method can effectively improve recall rate under the same accuracy rate.It can be used to detect various kinds of vehicles such as cars,vans,large trucks and so on in the actual road,and effectively overcomes the illumination,occlusion and viewpoint from environment.It shows good robustness.Furthermore,D-CNN with weighted multi-attribute strategy for extensive exploration of comprehensive vehicle attributes is proposed in this dissertation.The classical Multi-tasks Learning(MTL)method is combined with D-CNN.Unlike the traditional MTL method,the proposed method divides the tasks into main task and auxiliary tasks,and gives different weight values to them,that is aim to receive better model for main task.On the public vehicle dataset Comprehensive Cars(CompCars)and the vehicle data collected from Urumqi area of Xinjiang Province,the experiments based on the methods of single task,double task and multi-tasks are carried out.And the effectiveness of the method is also verified by the experiments.The experimental results show that the proposed method can realize the analysis about various internal attributes and external attributes of vehicle,and improve the accuracy rate by about 10%for the vehicle recognition and vehicle prediction on the CompCars dataset.So it has good applicable value and application prospects.