Another Vertical View: A Hierarchical Network for Heterogeneous Trajectory Prediction via Spectrums

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Trajectory prediction is expanding to complex, heterogeneous data forms like human skeletons and bounding boxes.
• Existing methods often overlook complex interactions within these diverse trajectory dimensions.
• The need for advanced AI techniques to handle multi-dimensional trajectory forecasting is growing.

Purpose of the Study:

• To extend trajectory prediction to handle heterogeneous trajectories of varying dimensions (M).
• To introduce a novel framework that models complex dimension-wise interactions.
• To improve forecasting accuracy for diverse trajectory types using AI.

Main Methods:

• Introduced trajectory dimensionality (M) to generalize the prediction task.
• Utilized Haar transform for capturing time-frequency properties across trajectory dimensions.
• Employed a bilinear structure to model time-frequency responses and dimension-wise interactions simultaneously.
• Developed a hierarchical forecasting approach using trajectory spectrums.

Main Results:

• The proposed model demonstrates superior performance on benchmark datasets (ETH-UCY, SDD, nuScenes, Human3.6M).
• Achieved state-of-the-art results in predicting heterogeneous trajectories, including 2D coordinates, bounding boxes, and 3D human skeletons.
• Effectively captured and leveraged dimension-wise interactions for improved prediction.

Conclusions:

• The generalized trajectory prediction framework successfully addresses heterogeneous data.
• The Haar transform and bilinear structure provide an effective way to model complex interactions.
• This work advances AI in trajectory forecasting for complex, real-world scenarios.