AFC-RNN: Adaptive Forgetting-Controlled Recurrent Neural Network for Pedestrian Trajectory Prediction

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian trajectory prediction is vital for computer vision tasks.
• Recurrent Neural Networks (RNNs) are commonly used for time series data like trajectories.
• Existing RNN methods inadequately model pedestrian memory and forgetting characteristics.

Purpose of the Study:

• To propose an Adaptive Forgetting-Controlled Recurrent Neural Network (AFC-RNN) for improved pedestrian trajectory prediction.
• To introduce an Adaptive Forgetting Controller (AFC) that adaptively manages historical data forgetting.
• To enhance the accuracy and reliability of trajectory prediction models.

Main Methods:

• Developed an Adaptive Forgetting Controller (AFC) using self-attention mechanisms to learn memory factors.
• Integrated the AFC into a Recurrent Neural Network (RNN) framework, creating AFC-RNN.
• Regulated the forgetting degree of historical trajectory features at each time step.

Main Results:

• AFC-RNN demonstrated superior performance compared to state-of-the-art methods on ETH, UCY, SDD, and NBA datasets.
• Extensive experiments and ablation studies validated the effectiveness of the proposed method.
• Mathematical analysis confirmed the advantages of adaptive forgetting over traditional RNN forgetting models.

Conclusions:

• The proposed AFC-RNN effectively models pedestrian trajectory prediction by adaptively controlling historical information forgetting.
• The novel Adaptive Forgetting Controller (AFC) significantly enhances prediction accuracy.
• This approach offers a more robust and accurate solution for trajectory prediction in computer vision.