(A) study on spatial relationship estimation model for moving objects in crowdsourcing environment

Abstract

Along with the advent of smart city, the intelligent transportation system is established to enhance the performance of traffic management in many aspects by providing innovative services. For example, there are several attempts such as gathering the data from CCTV systems to analyze and detecting traffic problems in the road environment. Since CCTV has the limitation on monitoring coverage compared to the urban scale, and in terms of cost, we try to utilize the crowdsourcing environment which does not require any installation cost and can gather the data from plenty of crowdsourcing observers through the roadside edge server to cover the vast area.

Practically, there are numerous studies conducted to collect and make use of crowdsourcing data. However, there are some problem issues that make analyzing the crowdsourcing data difficult. Since the data is collected from plenty of observers located at different locations and moving along the time, the viewpoints of data are also different over data. Therefore, direct comparison on spatial analysis by using two dimensional image contains distortion due to having different viewpoints among data, which can result in a critical error as the difference of viewpoints is larger. In addition, in crowdsourcing data, the information of each object in each data is not known to the other data. So re-identifying the object must be guaranteed in prior.

In this thesis, we propose a model to estimate the spatial relationships for moving objects in crowdsourcing environment. We focus on estimating the quantitative spatial relationship in distance between movable objects recorded as target analysis on crowdsourcing data. By estimating the quantitative distances between objects, several application services that utilize the spatial relationship between objects such as illegal parking detection can be emerged to improve the quality of life in the city. In the proposed model, we try to solve the problems that can occur by the different viewpoints among crowdsourcing data and estimate spatial information in data by using the properties that covisibility frequency represents the viewpoint invariant distance information between two objects and pixel-wise depth estimation can provide analyzing spatial information in unified three dimensional real world coordinate among data with different viewpoints. The model consists of three steps. In the first step, we intend to extract spatial information on the entire time frames of each video data into covisibility frequency vector so that the object-wise entire information observed among video data can be considered. By calculating the covisibility frequency for a multiple object pairs presented in several frames on video data taken with different viewpoints, the feature vectors representing spatial distance between objects are extracted from each video data, and by comparing the patterns of feature vectors representing the covisibility frequencies for matched object pairs among the data, we try to make it possible to compare the spatial relationship between objects among the data with different viewpoints. Although this method have an advantage that can cover wide range of target analyzing objects as it considers information of whole frames in video data, it lacks mitigating precise spatial analysis. To complement this limitation, at the second step and third step, we try to estimate depth information in pixel-wise from image data and extract feature vector that can represent the topological structure of objects' spatial information.

In the performance evaluation, we conducted the stepwise experiment using the simulated roadside data to verify the feasibility of the proposed method for analyzing the spatial relationship of the moving objects over crowdsourcing data. In the experiment of step 1, the extracted co-visibility pattern vectors of the data filmed at the same location with different viewpoints are quantitatively similar. For evaluation of step 2, the relative position map (RPM) estimator prove its ability to estimate the spatial relationship from single image with an accuracy of about 0.88 that results in high performance in the third experiment of RPM-based spatial relationship matching with a value of AUC is 0.8 for the case of 3 matching pairs. Through the experiments, we conclude that the proposed approaches can mitigate the existing problems of crowdsourcing data. Thus, it is applicable to analyze the spatial relationship among the movable objects in crowdsourcing data that can significantly contribute to diverse smart city services.