Multi-Level Feature Extraction and Classification for Lane Changing Behavior Prediction and POD-Based Evaluation
DOI: 10.3390/automation5030019
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the challenge of predicting lane-changing behavior (LCB) to enhance the safety and efficiency of advanced driver-assistance systems (ADAS) and autonomous vehicles. While machine learning (ML) algorithms are widely used for this purpose, their commercial adoption is limited by issues regarding robustness, reliability, and the inadequacy of standard evaluation metrics like receiver operating characteristic curves, which often ignore process parameters. The authors propose a methodology that combines multi-level feature extraction using deep autoencoders with ensemble classification techniques. Furthermore, they introduce a Probability of Detection (POD)-based evaluation framework to statistically assess classifier performance relative to the time remaining until a lane change, thereby accounting for process variability. The study utilizes data collected from a SCANeR™ driving simulator involving three human drivers performing lane changes and lane-keeping maneuvers on a simulated highway. The dataset includes 26 observation variables, such as vehicle velocity, distance to surrounding vehicles, steering angle, and pedal positions, recorded every 0.05 seconds. A deep autoencoder with four hidden layers was trained to extract multi-level features from this data. These features were then used to train four types of classifiers: artificial neural networks (ANNs), support vector machines (SVMs), hidden Markov models (HMMs), and random forests (RFs). Genetic algorithms were employed to optimize hyperparameters by minimizing a composite objective function balancing accuracy, detection rate, and false alarm rate. A "winner-take-all" ensemble strategy was applied, selecting the best-performing model among those trained on different autoencoder layers for each specific task. Performance was evaluated using the POD approach, which generates curves depicting the likelihood of correctly detecting a lane change as a function of time before the event. The evaluation focused on the $a_{90/95}$ metric, representing the time point at which the probability of detection reaches 90% with 95% confidence. Results indicated that ensemble ANNs were the most reliable classifiers, followed by SVMs. For lane changes to the left, the best predictions occurred 1.22 to 2.905 seconds before the event, depending on the driver. For lane changes to the right, predictions were made 2.823 to 4.355 seconds in advance. Notably, the optimal features for ANNs were predominantly found in the third layer of the autoencoder, demonstrating that high-level features do not always yield the best results. The method achieved low false alarm rates, with the best-case scenario predicting behavior approximately 6.5 seconds before the actual lane change. The significance of this work lies in its dual contribution to both model architecture and evaluation methodology. By integrating multi-level feature extraction with ensemble learning, the study demonstrates improved reliability in predicting human driving intentions earlier than previous methods. Additionally, the application of POD-based evaluation provides a more rigorous statistical assessment of ML models in dynamic contexts, explicitly accounting for the time-to-event parameter. This approach offers a robust framework for validating ADAS algorithms, potentially accelerating their deployment in commercial autonomous driving systems by ensuring higher confidence in safety-critical predictions.
Key finding
Ensemble artificial neural networks trained on multi-level features extracted from deep autoencoders provide the most reliable prediction of lane-changing behavior, capable of detecting the intent more than three seconds before the maneuver occurs.
Methodology
simulator
Sample size: 3
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 11 | 2026-06-06 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Theoretical Contribution: computational model