AD Behavioral Modeling: Lane Change Intent Prediction

This project implements a machine learning pipeline to predict vehicle lane change intent using the NGSIM (US-101 and I-80) trajectory datasets. The objective is to predict lane change intent as "Left" and "Right" before they occur using a 5-second prediction horizon. Data can be downloaded from here https://data.transportation.gov/Automobiles/Next-Generation-Simulation-NGSIM-Vehicle-Trajector/8ect-6jqj/about_data

Project Structure

• data/: Raw NGSIM trajectory data (US-101 and I-80).
• preprocess.py: Loads raw data, adds lane change intent labels with a 5s horizon, and applies physical lane-boundary constraints.
• models/:
  • rfclassifier.py: OOP-based Random Forest implementation including threshold optimization and model persistence.
  • ltsm.py: Sequential trajectory modeling (Next Steps).
• notebooks/: Playground area for jupyter notebooks and data exploration.
• results/: Logged runs containing Feature Importance plots, Confusion Matrices, and metrics_report.png.
• utils/:
  • data_prep.py: Handles vehicle-based splitting and sampling_keep_factor to prevent data leakage.
  • visualize.py: BEV (Bird’s Eye View) trajectory visualization.
• main.py: Orchestration script for the end-to-end pipeline.

Experimental Trials & Best Results

The model underwent iterative tuning to address the high noise floor in the NGSIM dataset and significant class imbalance.

Top Performance Metrics (Random Forest)

Class	Precision	Recall	F1-Score	Support
None (0)	0.98	0.99	0.98	1,698,261
Left (1)	0.49	0.43	0.46	38,805
Right (2)	0.35	0.26	0.30	14,462

• Optimal Thresholds: Left: 0.8052 | Right: 0.8442.
• Best Weights: {0: 1.0, 1: 3.0, 2: 10.0}.

Why "Left" outperforms "Right"?

In highway trajectory data, Left lane changes are typically more aggressive (overtaking), resulting in higher lateral velocity ($v_{lat}$) and clearer closing gaps. Right lane changes are often "lazier" (yielding or exiting), making them harder to distinguish from lane-drifting noise.

Key Insights & Feature Evolution

1. The Sensor Paradox: Raw, unclipped v_lat yielded better precision than filtered data, as the Random Forest utilized sensor "spikes" as early indicators of boundary crossing.
2. Contextual Logic: Adding Lane_ID and binary flags like can_go_right significantly improved precision by acting as physical logic gates.
3. Temporal Features: The addition of `v_lat_lag

Next Steps: LSTM Transition

To move beyond the current 0.46 (Left) and 0.30 (Right) F1-scores, the project is transitioning to a Long Short-Term Memory (LSTM) model to:

• Process trajectories as continuous temporal sequences rather than independent frames.
• Utilize hidden states to maintain long-term driving context.
• Improve "Interaction Awareness" by modeling how surrounding vehicle gaps influence intent over time.