Research On Vehicle Illegal Parking Detection Based On Lightweight Neural Networ

In recent years,with the significant improvement of people’s living standards,automobiles have become one of the most important modes of transportation.While the number of urban cars is growing rapidly,the demand for parking and the provision of parking lots are not compatible.Random parking and indiscriminate parking are common,which puts great pressure on traffic managers.In the face of the increasingly prominent problem of illegal parking,an accurate and fast illegal parking detection system is necessary,and it is of great significance to relieve the pressure of traffic police forces.In this paper,a deep learning algorithm is used to study the problem of vehicle parking violations.Acquire vehicle targets through object detection models.Then,the radiometric method and license plate recognition are used to determine the vehicle violation to realize the detection of vehicle violation in the area.Finally,the construction of the stop-violation detection system is completed on the embedded platform Jetson Xavier NX.This paper focuses on three main areas of work:(1)In order to solve the problem of large volume and poor real-time performance of the existing object detection algorithm,an improved method based on YOLOv5 s model is proposed.Firstly,the feature extraction structure of the model is optimized,the Bottleneck module is improved by using the Ghost convolution module,and the one-dimensional convolution in the C3 structure is replaced by the method of channel segmentation and channel superposition to reduce the amount of network parameters.At the same time,an improved attention mechanism is added to the three output paths of the backbone network to improve the accuracy of model detection.Finally,in the model training stage,mixed data is used to enhance the rich sample background of the dataset,and the EIOU loss function is introduced to improve the regression accuracy of the model.The improved model compresses the number of parameters by 35.79% and 36.16% respectively,ensuring accuracy and increasing the detection speed from 23 FPS to 28 FPS on embedded devices,and it has better performance on both datasets,reflecting the better generalization of the model.(2)Aiming at the task of vehicle violation detection,a method of violation discrimination based on radiometric method and license plate recognition is designed.When the number of vehicles in the area changes and is not 0,the LPRNet license plate recognition algorithm is used to record the detected license plate information and the corresponding time.This method does not need to use the target tracking algorithm to judge the driving state of the vehicle,and avoids the repeated extraction of vehicle features and the matching process of vehicles between frames.It can ensure the detection efficiency of the system.(3)The two algorithms of vehicle detection and violation detection are integrated to construct a complete vehicle violation detection system and implemented on the Jetson Xavier NX embedded platform.Tensor RT is used to further optimize the detection performance of the system,and the detection rate of the system is improved by simplifying the network structure of the model and reducing the calculation accuracy.Experiments show that the violation detection system designed in this paper can realize accurate and rapid detection of vehicle violation in the area,and has great application value.