Predictive queries on moving objects databases /
الاستعلامات التنبؤية على قاعدة بيانات الكائنات المتحركة
Mohammed Abdalla Mahmoud Youssif ; Supervised Hoda Mokhtar Omar Mokhtar , Neveen Elgamal
- Cairo : Mohammed Abdalla Mahmoud Youssif , 2020
- 80 Leaves : charts ; 30cm
Thesis (Ph.D.) - Cairo University - Faculty of Computers and Artificial Intelligenc - Department of Information Systems
Future trajectory prediction for moving objects, e.g., vehicles, has a significant impact on many location-based services such as location-aware search, traffic management, mobile advertising, and travel guidance. The existing techniques which predict the future path(s) of moving objects depend mainly on their motions history to perform the prediction process. As a result, these techniques fail when moving objects history is unavailable. This thesis aims to present efficient solutions for predicting the trajectories of moving objects without relying on their past trajectories. The proposed solutions include - (1) SimilarMove: a similarity-based prediction system for moving object future path, (2) DeepMotions: a deep learning system for moving object future path prediction, and (3) SAM: a spatial attention model for future trajectory prediction.The main idea of SimilarMove is obtaining the future paths of the query moving object in terms of other objects currently moving similar to the query object. After that, SimilarMove employs a Hidden Markov Model that receive these similar trajectories as an input and generates the possible future paths with their related probabilities as an output.The DeepMotions extracts the latent motion patterns from K nearest neighbor similar objects moving like the query moving object. Then, a Bi-directional recurrent deep-learning model is built based on these extracted motions and generate predictions. The main idea of SAM is to generate predictions by not scanning the whole input trajectory sequence but, focuses only on the significant positions of the input trajectory sequences to produce the output. This allows the internal representation of input trajectories to be refined based on the relevant information from the query object. Then, by gathering relevant information into the final representation, only the necessary information is provided to predict the final answer of the query object