Research On Algorithm For Recognition Of Vehicle Types Under Multi-Task Learning

The vehicle identification system has important significance for the intelligent transportation system,which can effectively monitor suspicious vehicles,count traffic flow information,and so on.Although the algorithm based on shallow machine learning is efficient,it needs to design different features for different scenarios.In practical applications,the scene is complex and changeable.This method obviously cannot meet the actual requirements.The emergence of deep learning brings new ideas to the vehicle identification system in complex environmental scenarios.It can use massive training data to automatically learn more general features.However,its basic theory is still not perfect,and its effectiveness needs to be verified by experiments.In practical applications,due to the difficulty of data collection and labeling,there is often a lack of sufficient sample data.Moreover,most convolutional neural network models are classified and detected for general objects,and need to be improved in specific application scenarios.According to the actual source of the problem,the problems of vehicle detection and vehicle identification under the complex bayonet scene are analyzed.For the single detection of the general target detector in complex scenes,the problem of missed detection and false detection will occur.The vehicle detection data set under different weather scenes is set up,and the K-means clustering algorithm is used to analyze the bayonet vehicle detection data set.Distribution characteristics,based on the Single Shot Multibox Detector,set the region candidate frame with appropriate scale and aspect ratio,delete the redundant region candidate box,and improve the general detection algorithm.For the problem of model identification,the data set is small,the feature difference is small,and the environment is complex and changeable,which leads to over-fitting.From the perspective of data expansion,migration learning strategy and design model,the problem of less data volume is alleviated.Data expansion is based on existing data samples,and the data samples are expanded by certain rules;the migration learning strategy makes full use of existing data samples;the design model is to design a suitable number of Residual Network models,reduce the model capacity and avoid excessive sampling noise in the model,forces the model to learn more distinguishing features while speeding up the model's operating rate.Finally,based on the idea of multi-task learning,combined with classification learning and metric learning,fully explore the supervision information between tags and tags,and further constrain the learning of parameters.The experimental results show that for the specific data set,the regional candidate frame is set near the clustering center of the training data,which reduces the noise interference,and deletes the redundant candidate frame,can learn more general features,and the position regression is also more Easy,improve the performance of vehicle detection.In the case where the data expansion strategy and the model structure are consistent,migration learning by using the well-trained good model in the data with certain similarities,the knowledge available therein is migrated,and the small sample data set is fine-tuned to greatly improve the model.The performance of the identification,in addition,learning the local information between the sample tag information and the sample tag at the same time,can further prevent the network from over-fitting when the amount of data is small,and improve the performance of the vehicle identification.