New AI Technique Predicts Object Movements Across Multiple Surveillance Cameras

Researchers have developed a new approach to track and predict the movements of objects across multiple camera views simultaneously—a challenge that matters for surveillance systems, traffic monitoring, and other applications that rely on networked cameras but has been largely overlooked by existing forecasting methods.

The Problem with Single-Camera Forecasting#

Current trajectory prediction systems typically work within the constraints of a single camera's field of view, making them unsuitable for real-world multi-camera deployments. This limitation means they cannot predict long-term movements or seamlessly follow an object as it transitions between different camera views. The researchers address this gap by developing a framework that processes observations from several viewpoints at once.

Introducing Trajectory Tensors#

The core innovation is a concept called trajectory tensors—a new way to represent and encode movement paths as objects travel across multiple camera perspectives. This technique also captures the associated uncertainties in the predictions. Rather than simply tracking pixel coordinates, trajectory tensors allow the system to reason about both the spatial and temporal dimensions of movement across an entire camera network.

The "Which-When-Where" Approach#

The proposed framework answers three critical questions:

• Which camera(s) will the object appear in next?
• When will it arrive in those views?
• Where exactly within each camera frame will it be located?

The researchers built several encoder-decoder models capable of processing trajectory tensors and tested them on a custom dataset containing 600 hours of video from 15 camera views. Their experiments show that models based on trajectory tensors substantially outperform methods that rely on traditional coordinate-based tracking, as well as single-camera forecasting techniques adapted for multi-camera settings.

The work is available on arXiv.