Research On Roadside Parking Space Monitoring Based On Machine Vision

Building an intelligent and convenient parking environment is of great significance for the development of Intelligent Cities.On-street parking spaces occupy a very important position in the urban parking ecology due to its wide distribution.Nowadays,on-street parking spaces are mostly manually monitored,and their registration efficiency is truly not high.The charging procedure is cumbersome,resulting in a poor parking experience.Thus,improving the intelligent monitoring and management level of parking spaces is a significant means to solve these problems.After analyzing the current situation and difficulties of parking spaces monitoring,this paper presents an imagery-based framework to infer parking space status by generating 3D bounding box of the vehicle based on machine vision.It focuses on vehicle detection in high-occlusion environments,3D attitude estimation of vehicles,top-view construction,and multinomial logistic regression model for spaces status refinement.The main researches of this work are summarized as follows:(1)The research of vehicle proposal generation under high occlusion environment.We integrate Faster R-CNN as benchmark,and optimize its Region Proposal Network(RPN)as part-aware RPN,where the vehicle is divided into several independent parts and the characteristics of different parts are extracted separately,generating proposals for different parts.Then,different parts belong to the same vehicle are encoded and reconfigured into a compositional entire proposal through a Part Affinity Fields,allowing the model to generate integral candidates and mitigate the impact of occlusion challenge to the utmost extent.At the same time,we propose a part-aware Non-maximum suppression based on the part proposals,which is utilized to optimize the dilemma of traditional non-maximum suppression in occlusion environment.(2)Concerning the problem of multi-scale on vehicle detection,a specially designed convolutional neural network based on Res Net and feature pyramid network is proposed to overcome challenges from partial visualization and occlusion.It predicts 3D box candidates on multi-scale feature maps with five different 3D anchors,which generated by clustering diverse scales of ground truth box according to different vehicle templates in the source data set.(3)Parking space top view is constructed and parking space status inference is completed through a Multinomial Logistic Regression Model(MLRM).A top view of the parking space is constructed through the projection matrix to directly observe the pose of the parking vehicle.Considering that the status inference will be disturbed by non-standard parking behavior and easily influenced by vehicles in adjacent parking spaces.In the procedure of space status inference,MLRM is trained to model this process and complete inference by combining three adjacent parking spaces into one unit.This guarantees that the outputs won’t be interfered by adjacent parked vehicles.(4)Through the comparative experiments on KITTI and our dataset,the proposed 2D vehicle detector constructed by part-aware RPN is able to conquer most challenges such as occlusion and truncation.Compared with the state-of-the-art vehicle detection approaches,the proposed method indeed improves the performance of accuracy especially when vehicles are heavily occluded.At the same time,this paper also validates the performance of 3D vehicle detection pipeline generated by combining 2D proposals and 3D anchors.Several experiments in real environments were performed with this system demonstrate their functionality compliance and relatively good performance,comparable to that of the most recent works in the field.