Figure 1. Crowdsourcing based Road Environment Report & Domain Event Data Management Platform
As the various and many mobile devices come into wide use, it becomes possible to improve the quality of life for citizens by providing the smarter service through being reported the local information of each mobile and using them. In the crowdsourcing based road environment report & Domain Event Data Management Platform, the illegal parking re-identification to solve the traffic congestion smartly and validation reliability of the target event & classification duplicate events to upload crowdsourcing event uploaded by various citizens to the event sharing map service.
1. Semantic aware Road Environment Data Re-identification
Figure 2. Multi-level Image Feature Extraction for Object Re-identification
– To verify that images from two different view point is spatially identical, analyze the image in three level(Point-level, Scene-level, Object-level)
– In order to match the texture of images, extract the point-level feature by using Scale-Invariant Feature Transform(SIFT) algorithm.
– For the scene-level and object-level feature matching, we will develop deep learning model to analyze the semantic information finally.
– By analyzing the point-level matching similarity and semantic information, comprehensive judgment for the illegal parking re-identification is conducted.
2. Vehicle Movement Classification around Road
– Verify whether the specific vehicle, spotted at a place through time-varying multiple crowdsourcing inputs, has moved or not.
– Each Crowdsourcing input videos has different viewpoint and time on the same place. Such differences are first considered for further detecting illegal parking situation.
– Viewpoint difference is managed via image viewpoint matching transformation with SIFT image feature extraction and matching procedure.
– Deep Neural Network(Residual Net-50 based Mask R-CNN) based image segmentation technique is adopted for accurate vehicle detection and vehicle movement detection between two different crowdsourcing input videos.
3. Scene Analysis based Context-aware Domain Event Management
- – Among the hundreds of Domain Event data generated in smart city life, when interest information of individual citizens gathers and becomes a local concern, this is defined as a smart city event.
– While crowdsourcing based collected data has the advantage of collecting a large amount of information by multiple participants, some participants often intentionally distort information or enter missing information by mistake, so it is necessary to verify the data is reliable. And, when multiple participants upload information on the same event, a data management function is needed to determine and classify duplicate events.
– To solve crowdsourcing data problems, We devised an algorithm to analyze crowdsourcing event data. This algorithm applies element techniques such as image scene analysis, spatio-temporal / domain clustering, and description comparison analysis based on the NLP technique LDA topic modeling algorithm.
Figure 3. Procedure of Crowdsourcing based Smart City Event Reliability Validation and Duplicate Event Group Classification
– When the target event is uploaded, it is checked if there is image data. If there is image data, the situation of image scene is analyzed through multiple object detection algorithm and context-aware based domain inference deep learning model. The results are reflected when verifying event reliability and determining if there are duplicates.
– Then, we extract the keywords from the text document and apply LDA algorithm-based topic modeling to learn which topics belong to the entire event and classify event groups to be duplicates of target events.
– Topic modeling is performed several times in proportion to the total number of events, classifying ① event group belonging to the topic similar to the target event. In order to classify ② event group, first, similar event groups are classified by domain information identical to the target event, event start date within 5 days, and K-means clustering of location information in all events. Subsequently, topic modeling for similar event groups is performed to classify ② event group. By comparing and analyzing ① event group and ② event group, we devised the logic of event reliability measurement and duplicate event determination.