Research And Implementation Of Vehicle Detection Method Based On Convolution Neural Network

With the rapid development of social economy,the number of vehicle owners in China keeps growing.The illegal and criminal events of the vehicle are rising year by year.However,the existing traffic control technology is backward and the degree of intelligence is low.Although the electronic police system can,to a certain extent,curb all kinds of vehicle-related crimes,most systems have some problems,such as the shortage of vehicle information,the low accuracy of vehicle information retrieval by vehicle license plate recognition,the high degree of manual intervention and so on.The vehicle detection method can provide more vehicle information to match with the vehicle information database,and confirm the vehicle “identity” by introducing the vehicle type information to improve the accuracy of vehicle information retrieval.In order to achieve better vehicle detection,we proposes two vehicle detection models based on convolutional neural network by improving the YOLOv2 object detection model,and designs and implements a vehicle detection micro service based on improved model.The main work of this thesis includes:(1)Explaining the significance of vehicle detection research,analyzing the advantages and disadvantages of current vehicle detection algorithms combining with convolutional neural network and object detection model.(2)Elaborating the design ideas of anchor boxes used for positioning,activation function and loss function in the model of vehicle detection.The Model-A is designed by combining feature fusion strategy,and the Model-B is designed by removing the high-level repeated convolution operation strategy.(3)The test result on the BIT-Vehicle data set shows that the Model-A detection accuracy reaches 94.16% and the Model-B detection accuracy reaches 94.78%.It proves that the two improved models are both suitable for the vehicle detection task under the road monitoring and the detection results are similar.To further compare the generalization capabilities of the two improved models,the two improved models are tested on the CompCars data set.The result on the CompCars migration test data set shows that the Model-B is better,indicating that it is more suitable for the transfer learning research of vehicle detection tasks,and validates the effectiveness of removing the high-level repeated convolution operation strategy.In order to analyze the improved model in depth,we take the Model-B as an example to visualize the network information,and more intuitively shows its better vehicle feature extraction capability.(4)Based on the comprehensive experimental results,we select Model-B as the theoretical basis,design and implement it as a simple and highly scalable vehicle detection micro service,and display the main functional interface.The work of this thesis provides an efficient vehicle detection method,which combines the micro service architecture to implement a vehicle detection micro-service.It can serve more customers and collect more vehicle data to lay the foundation for reinforcement learning in future.