Research On Vehicle Detection And Vehicle Logo Recognition Based On Deep Learning

The rapid growth of car ownership in recent years has brought enormous pressure to the development of urban intelligence and brought problems such as traffic congestion,frequent accidents and air pollution.The traditional traffic management methods cannot deal with these problems.Intelligent Transportation System has become an important technical means to solve vehicle speed control,traffic control,traffic violation detection,road toll and other related problems by using traffic monitoring video.The detection of vehicle objects and their attributes is an important part of the development of the Intelligent Transportation System.It is very difficult to integrate the task of object localization and classification of attributes.Moreover,based on the traditional target detection technology,the detection accuracy is low and the speed is slow.Therefore,the detection technology based on deep learning has developed rapidly and achieved good results in the field of vehicle detection.However,the current public datasets of vehicle detection contain a large number of samples with simple background and a small number of targets to be detected in a single image,which cannot meet the needs of real traffic monitoring and detection.And the model with high detection accuracy has a slow detection speed,which cannot meet the requirements of real-time detection.In addition,the researchable samples for the detection of vehicle logo attributes are only the local position images of the vehicle logo,and its research content cannot be applied to the video surveillance detection task.In order to solve the above problems,this thesis conducts research from three perspectives of reconstructing data sets,improving detection accuracy and speed.The main contents are as follows:(1)Construction of the dataset.Select local highway traffic checkpoint monitoring videos for sample extraction.The samples should be intercepted by frames from the monitoring videos taken under different lighting conditions and different weather,such as day,night,sunny day and rainy day;and select the images with learning value for labeling.In this process,it is necessary to avoid the imbalance of positive and negative samples caused by the labeling work.Then select the samples with clear front face and tail range from the vehicle image for the next step of vehicle logo recognition.(2)Research on multi-class vehicle detection algorithms.The CenterNet algorithm is used to detect 8 common types of vehicles.In order to meet the real-time detection requirements,the standard convolution in the original model is replaced by a depthwise separable convolution,and the amount of model parameters is reduced under the premise of less impact on the detection accuracy to speed up the detection.Then,the attention mechanism strategy is introduced into the model backbone extraction network residual module to enhance the model’s extraction of effective features of the detected image;GHM is used to optimize the original size regression loss function to improve the accuracy of the detection target regression size information.Training and learning are carried out on the self-made multi-category vehicle data set.The experimental results show that the mAP and FPS reach 95.23%and 50 frames per second,respectively,which can meet the requirements of real-time detection of surveillance video under the premise of improved detection accuracy.(3)Research on vehicle logo recognition algorithm.The SSD algorithm is used for vehicle logo recognition.The SSD is improved from three perspectives of network structure,feature fusion,and data normalization to improve its detection effect on small targets.The residual module is introduced to replace the original backbone network,and the feature fusion strategy is optimized.Select Group Normalization to process data.The experiment on the self-made car logo dataset shows that mAP reaches 94.12%,which is 2.08%higher than the original model.