(A) study on adaptive object re-identification scheme using RoI-based SIFT-CNN model in crowdsourcing environment

Abstract

Along with advancement of the IoT (Internet of Things), demands for smart cities are increasing to efficiently manage resources and assets of the city. Among these smart city areas, the surveillance systems account for the large portion and video surveillance systems are the largest market size in the surveillance area. The surveillance systems used for object re-identification are popular using CCTV solutions. However, this environment is expensive to implement applications. Therefore, a new data collection approach is needed. The crowdsourcing environment is considered an approach of re-identification as a cost-effective manner. In crowdsourcing environment, various observers send the road data to edge server. Observers can easily move all road, so edge server collects all of road data not by collecting limited road area such as CCTV system. Also, observers have the camera to record a road, crowdsourcing environment costs little to install. This is why we implement a surveillance system at a low cost instead of CCTV. Although there are techniques for re-identification or identification objects in the past, traditional techniques are difficult to apply them to the smart city environment because the means of collecting data are only considered for CCTV, not for crowdsourcing.

In this thesis, we proposed object re-identification scheme in spatio-temporal and crowdsourcing environment. We consider Region of Interest (RoI) to improve the accuracy of the object re-identification. The object has a common characteristic depending on the observation angle, and the area with each common characteristic is set to the same RoI. In this thesis we divided RoI into 5 regions, which are experimentally obtained. Thus we propose the object re-identification technique by considering the RoI of an object when we process object re-identification in a crowdsourcing environment. When we identify an object with RoI, we compare identical characteristics in the same RoI. So we improve the accuracy of object re-identification. CNN is used to extract the object from scene image in data of observers to obtain the same class of the objects and viewpoint of the object. When matching objects, the Euclidean distance is calculated to determine the target object. The average distance of object is minimum. To validate the proposed technique, we compared the SIFT-CNN feature matching approach which does not consider RoI our proposed model, when the object re-identification in an environment in which video was taken by different observers for nine scenes is transmitted on roads in KAIST campus. The experiment results showed that the overall mAP of the RoI considering technique increased by 0.08 and the accuracy of the RoI considering technique increased by 2.1% compared to a technique that is not considered RoI, finally we improved the accuracy of the object re-identification.