Friday 31 January 2025
A team of researchers has developed a new system for tracking objects in multiple camera views, which could have significant implications for applications such as autonomous vehicles and surveillance systems.
The system, called BEV-SUSHI, uses a combination of computer vision techniques and machine learning algorithms to track objects across different cameras. It starts by detecting objects in individual camera views using a deep learning-based method called BEVFormer. This method is trained on synthetic data generated from 3D models of the environment.
Once the objects have been detected, the system then uses a process called re-identification (ReID) to match them across different cameras. ReID works by extracting features from images of the objects and comparing them to identify matches.
To improve the accuracy of the ReID process, the researchers developed a new algorithm that associates 2D detections in individual camera views with projected 3D detections from BEVFormer. This helps to reduce errors caused by noisy backgrounds or other objects in the image crops.
The system was tested on two datasets: AICity’24, which is a synthetic dataset generated from 3D models of environments, and WildTrack, which is a real-world dataset collected from multiple cameras.
Results showed that BEV-SUSHI achieved impressive tracking performance on both datasets, with mAP scores of over 90% on the AICity’24 dataset and 77.18% on the WildTrack dataset. The system was also able to track objects across long sequences of images, demonstrating its ability to handle complex scenarios.
The researchers believe that BEV-SUSHI could have significant implications for applications such as autonomous vehicles, where accurate tracking of objects is crucial for safe navigation. It could also be used in surveillance systems, where it could help to improve the accuracy and efficiency of object detection and tracking.
Overall, the development of BEV-SUSHI represents a major step forward in the field of computer vision and machine learning, with significant potential applications in a range of fields.
Cite this article: “Multicamera Object Tracking System Achieves High Accuracy”, The Science Archive, 2025.
Object Tracking, Multiple Camera Views, Bev-Sushi, Computer Vision, Machine Learning, Autonomous Vehicles, Surveillance Systems, Re-Identification, 3D Models, Deep Learning
