Suspicious Object Monitoring System Based on Multiview Model from Smart-camera

The context of reporting on multiple cameras some-what overlaps. The paper solves three problems. The first is to detect pedestrian objects on each camera. The second is to capture and assign IDs to each object on each camera. Finally, unify IDs for objects on all cameras. For object recognition, through the survey and testing process, we chose the YOLOv4 model - a popular model used in object detection. The model was trained with new data, and matched to the pedestrian scene, achieving an accuracy of 95.6%. Continuing with the object tracking operation, we use the DeepSORT algorithm (tracking by detect algorithm) with the main input being the output of YOLOv4 achieving a MOTA of 40.85%. For the final problem, we bring all the objects from the cameras to a common coordinate system, then use the extracted positions and features of the objects to synchronize the IDs across all the cameras and achieve an IDF of 80.2.