Research On Vehicle Counting Method Based On Attention Mechanism And State Detection

Vehicle detection and counting are of great significance to intelligent road monitoring.Accurate and real-time monitoring of traffic data on traffic roads is a hot issue in the realization of intelligent transportation systems.By automatically counting the entering or leaving of different types of vehicles at each time period,it has important reference value for realizing intelligent traffic scheduling and optimization of road conditions.However,problems such as low accuracy of vehicle counting and large consumption of computing resources are common.In order to balance the accuracy and speed of counting,this paper proposes a saliency enhancement module based on the attention mechanism and integrates it into the Mobile Net V1+SSD（Single Shot Detector）model as vehicle detection network.In view of the excessive consumption of computing resources in the tracking process,this paper proposes a filtering mechanism based on the vehicle state,which reduces the number of tracking in the vehicle counting process.The main work is as follows:（1）A multi-objective dataset based on the vehicle state is produced.The main work is divided into several steps such as collecting video,cropping video pictures,filtering pictures,marking pictures and data enhancement.By introducing the concept of vehicle state,the auxiliary position recognition line L is added to the vehicle picture,and then when the data set is labeled,it will be divided into four states according to the relative position of L and the vehicle:the vehicle is in front of L（up state）,The vehicle is in L（mid state）,the vehicle is behind L（down state）,and the vehicle is far away from L（far state）.A data set of 24,794images is marked.（2）In order to improve the detection accuracy of the vehicle state,an attention-based saliency enhancement（ASE）module is proposed.This module uses the attention and receptive fields of the deeper feature maps to enhance the shallower feature maps,and then uses the maximum operation method for the enhanced feature maps and the original feature maps to obtain significant enhancement features.Then use the Mobile Net V1+SSD network integrated with 3 ASE modules for detection,so that the vehicle detection reaches a certain balance in detection accuracy and speed.（3）A filtering mechanism based on vehicle status is proposed,and a vehicle counting method based on status filtering is implemented.In the process of vehicle counting,the purpose of compressing the vehicle counting space is achieved by excluding all vehicles in the far state,thereby effectively reducing the number of tracking and related calculations.The detection network integrates SSD with 3 ASE modules as vehicle detectors(SSD-ASE_1,2,3).After ablation experiments and performance evaluations with several state-of-the-art methods.With a slight increase in the amount of calculation,the proposed network obtains a higher detection accuracy（m AP=95.94%）.The vehicle counting experiment first compared several different counting methods in recent years,and then compared the proposed state-based vehicle counting method with the traditional location-based vehicle counting method.Experimental results show that the proposed method has certain advantages in counting accuracy,and the tracking times of traditional counting methods are reduced by about 60%on average.