Enhanced Fusion of Spatial and Temporal Features for Accurate Vehicle Trajectory Prediction (GP-STFF)

---

📌 Abstract

Accurate and resource-efficient trajectory prediction for surrounding vehicles is essential for autonomous driving. Existing Graph Convolutional Network (GCN) and attention-based methods often consume excessive computational resources by processing all historical trajectories.

To address this challenge, we introduce GP-STFF (Graph Positional and selected Spatial-Temporal Feature Fusion)[cite: 1]:

1. Positional Graph (GP): Combines relative distance graph structures with learnable Position Encoding (PE) to capture both relative interactions and absolute spatial positions[cite: 1].
2. Selected Spatial-Temporal Feature Fusion (selected-STFF): Selects spatial interactions at specific key time steps (e.g., $t=1$) and fuses them with temporal features via a Multi-Head Self-Attention mechanism, drastically reducing computational overhead while retaining accuracy[cite: 1].
3. Performance: Achieves a 9% reduction in average RMSE compared to state-of-the-art methods on the public NGSIM dataset[cite: 1].

---

🏗️ Model Architecture

The GP-STFF model consists of:

• Spatial Dimension Module: Constructs the relative distance relationship graph $G_w$ and adds learnable Position Encoding $PE$ to build $G_{GP} = G_w + PE$[cite: 1].
• Temporal Dimension Module & Selected-STFF: Processes historical trajectory $X$ of the target vehicle and fuses it with selective $G_{GP}$ using an LSTM and Multi-Head Self-Attention mechanism[cite: 1].
• Feature Aggregation & Decoder: Concatenates spatial-temporal features with surrounding vehicle trajectory representations ($F_n$), applies an attention bottleneck, and outputs future trajectory points ($Y$) via an LSTM decoder[cite: 1].

---

🌟 Key Contributions

• Efficient Spatial-Temporal Integration: Seamlessly integrates spatial factors directly into time-series trajectory learning[cite: 1].
• Positional Graph Representation: Captures both relative inter-vehicle distances and absolute positional importance through learnable position encodings[cite: 1].
• Interpretability: Provides qualitative insights into driver habits and attention distributions across different vehicle types (e.g., Automobiles vs. Heavy Trucks)[cite: 1].

---

📊 Benchmark Results

Evaluated on the NGSIM highway dataset (US-101 and I-80) using Root-Mean-Square Error (RMSE) across 5 prediction time steps (1.0 second horizon at 0.2s/step)[cite: 1]:

Model	1st Step	2nd Step	3rd Step	4th Step	5th Step
Naive LSTM[cite: 1]	0.1012	0.2093	0.3384	0.4830	0.6406
CS-LSTM (CVPRW '18)[cite: 1]	0.1029	0.2023	0.3146	0.4364	0.5674
WSiP (AAAI '23)[cite: 1]	0.1030	0.1998	0.3114	0.4340	0.5631
STA-LSTM (IEEE T-ITS '22)[cite: 1]	0.0995	0.2002	0.3130	0.4348	0.5615
STA-LSTM + Graph[cite: 1]	0.0827	0.1825	0.2965	0.4183	0.5462
GP-STFF (Ours)[cite: 1]	0.0769	0.1727	0.2851	0.4057	0.5311

---

🔄 Trajectory Prediction Pipeline

The overall execution pipeline of the GP-STFF trajectory prediction framework follows a systematic, end-to-end workflow:

---

Step-by-Step Workflow

0. Vehicle ID Tracking & Sequence Assignment

To ensure consistent temporal tracking and multi-agent spatial interaction modeling, each vehicle within the ROI (Region of Interest) is assigned a unique identifier sequence (ID tracklet) across all consecutive observation frames.

Figure 0: Unique ID sequence assignment for target and surrounding traffic agents across temporal frames.

---

1. Historical Trajectory Data Collection

First, historical trajectory data of the ego vehicle (target vehicle) are extracted across the highway segment over the predefined observation horizon ($T_{obs}$).

Figure 1: Extraction of historical trajectory points for the target vehicle on the road section.

---

2. Surroundings Perception & Multi-Vehicle Radar Sensing

On-board sensors (e.g., radar, LiDAR) collect spatial context and historical motion trajectories from surrounding traffic agents within the local grid interaction area.

Figure 2: Perception field capturing surrounding vehicles' historical movement patterns via vehicle radar.

---

3. Spatial-Temporal Feature Fusion & Model Inference

The collected spatial-temporal sequences are processed using the trained GP-STFF architecture[cite: 1]. The positional graph ($G_{GP}$) combines topological graph distances with learnable position encodings ($PE$), while the selected Spatial-Temporal Feature Fusion (selected-STFF) module extracts cross-modal interactions to forecast future states over the prediction horizon ($T_{pred}$)[cite: 1].

Figure 3: Structural diagram of the trained GP-STFF prediction model.

---

4. Future Trajectory Forecasting & Visualization

The decoder outputs sequence-level future trajectory coordinates ($\hat{Y}$), which are rendered onto the spatial domain to demonstrate lane-keeping, lane-changing, and speed variation behaviors.

Figure 4: Visualized trajectory prediction outputs for target and surrounding traffic agents.

---

5. Density Estimation & Probability Distribution Analysis

To quantify prediction uncertainty and multi-modal trajectory distributions, fine-grained spatio-temporal features are analyzed using spatial heatmaps and positional probability density point maps.

Figure 5: Spatial trajectory uncertainty heatmap.

Figure 6: Predicted trajectory probability density distribution map.

---

🚀 Getting Started

Environment Setup

# Clone repository
git clone [https://github.com/victor1469/GP-STFF.git](https://github.com/victor1469/GP-STFF.git)
cd GP-STFF

# Install requirements
pip install -r requirements.txt