Robust Hybrid Approach of Vision-Based Tracking and Radio-Based Identification and Localization for 3D Tracking of Multiple Construction Workers

Abstract

Information of construction entity identity and real-time location reveals where specific construction resources are at any given time and thus is a critical prerequisite to the context-aware jobsite safety management. Most existing studies use either a vision-based or radio-based approach to automatically track construction entities, which, however, are bounded by the limitations of the applied technology. Vision-based tracking can achieve high localization accuracy but suffers from identity (ID) switch and fragmentation errors when multiple workers are in close proximity or occluded. In contrast, radio-based tracking is reliable in object detection and identification but less accurate in localization. This study proposes a hybrid framework that fuses results obtained from vision-based tracking and radio-based identification and localization for the 3D tracking of multiple construction workers. Compared to traditional fusion approaches that directly fuse locations extracted from these two approaches, the proposed method treats vision-based tracking as the main source to extract the object trajectory. Radio-based identification and localization results are used as a supplementary source to augment anonymous visual tracks with identity information and correct errors (e.g., false positives) in vision-based object detection, resulting in ID-linked 3D trajectories. In addition, a searching algorithm is introduced to recover possible missed detections in one camera view from the corresponding observations in the other view by applying a sliding window to search for regions with the most similar appearance along epipolar line. The newly created method has been validated using two indoor experiments. The results show that the new approach for fusing vision- and radio-based results increases the overall accuracy from 88% and 87% to 95% and 90%, compared to using a vision-based approach alone. The integration of radio-based identification is much more robust than using a vision system alone, as it allows the recovery of the same entity ID after the trajectory is fragmented and results in fewer fragmentations that last longer than 0.2 s.