Predicting Motion Direction and Person re- identification across surveillance Cameras Network using (LSTM)

cameras surveillance systems are develop and spread in our life because it used of various fields as monitoring the interested places, buildings and traffic. Prediction of moving object in scene still challenge as the object when change his appearance or may be occluded. This paper presents an efficient proposed system for tracking and predicting the motion direction of a person by combining deep learning and topology of a camera network in non-overlapping multi-camera surveillance. The proposed system is divided into three stages: firstly, extracted suitable features of person in single camera using YOLOv3 algorithm for target detection; secondly, tracking the target based on deep sort and Kalman filter to association observations data for unique person across cameras; finally, person re-identification when a person moving across adj acent cameras by using interested information for target with Long Short-Term Memory (LSTM) network and spatial information from cameras network topology to esti-mate the direction of target when move between cameras that will improve the tracking process and reduce the processing and traffic on bandwidth.