Multi-features taxi destination prediction with frequency domain processing.

Area of Science:

• Transportation Science
• Computer Science
• Data Science

Background:

• Traditional taxi prediction models spatial points, missing 2D relationships.
• Trajectory images can suffer from noise and sparsity.
• Existing methods struggle to represent complex spatial correlations effectively.

Purpose of the Study:

• To propose a novel method for taxi destination prediction that addresses noise and sparsity in trajectory images.
• To enhance prediction accuracy by incorporating frequency domain processing and multi-feature integration.
• To improve the representation of spatial relationships in taxi trajectories.

Main Methods:

• Transforming spatial taxi trajectories into frequency-domain representations using Fast Fourier Transform (FFT).
• Employing Convolutional Neural Networks (CNN) for deep feature extraction from processed trajectory images.
• Utilizing Recurrent Neural Networks (RNN) for destination prediction, integrating extracted features with trajectory metadata and data.

Main Results:

• The proposed multi-features taxi destination prediction with frequency domain processing (MTDP-FD) method significantly reduces noise and sparsity.
• MTDP-FD achieved a 0.14km reduction in average distance error compared to existing methods.
• The GTOHL combination of data and features proved most effective in enhancing prediction accuracy.

Conclusions:

• Frequency domain processing is effective in denoising taxi trajectory images for improved destination prediction.
• Integrating deep features from CNNs with RNNs and multi-modal data enhances prediction performance.
• The MTDP-FD method offers a more accurate and robust approach to taxi destination prediction.