DriveNet: A deep learning framework with attention mechanism for early driving maneuver prediction
DOI: 10.11591/ijai.v14.i1.pp44-53
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 introduces DriveNet, a deep learning framework designed for the early prediction of driving maneuvers to enhance road safety and support Advanced Driver Assistance Systems (ADAS). The research is motivated by the high prevalence of accidents caused by inappropriate driving behaviors and the critical need for systems that can anticipate driver intentions with sufficient lead time for human reaction. Existing methods often neglect the specific time gap between prediction and execution, which is vital for effective driver assistance. DriveNet addresses this by integrating multimodal data—specifically driver behavior and environmental context—to predict maneuvers up to four seconds in advance. The proposed architecture combines spatial and temporal feature extraction techniques. It utilizes a VGG19 convolutional neural network to extract spatial features from driver-facing video frames and the OpenFace framework to derive 54 face-based features, including eye gaze and head pose. Environmental data, such as vehicle speed, empty lanes, and road artifacts, are also processed. These inputs are fed into a bidirectional Long Short-Term Memory (BiLSTM) network enhanced with an attention mechanism to capture long-term temporal dependencies. The system was evaluated using the publicly available Brain4Cars dataset, which contains naturalistic driving data from 10 drivers. Due to the limited sample size (594 accessible maneuvers), the authors employed data augmentation via sliding windows. Experiments were conducted using 5-fold cross-validation, assessing performance across five time-to-maneuver intervals (0 to 4 seconds) for five maneuver classes: left/right lane changes, left/right turns, and straight driving. DriveNet achieved state-of-the-art performance, particularly in early prediction scenarios. For predictions made four seconds before a maneuver, the model attained an accuracy of 91.24%, precision of 90.63%, and recall of 90.32%. The system demonstrated superior efficacy in predicting turns (accuracy >97%) compared to lane changes (accuracy ~88–89%). When aggregating predictions from current and previous sequences, performance improved further, reaching 96.23% accuracy at the moment of maneuver execution. Comparative analysis showed DriveNet outperformed previous methods like DBRNN and CF-RNN, especially regarding the lead time provided to drivers. While STA-Net achieved slightly higher metrics, it did so with zero seconds of lead time, whereas DriveNet provides a critical four-second window for driver reaction. The significance of this work lies in its ability to provide actionable early warnings for ADAS, potentially mitigating accidents caused by illegal or dangerous maneuvers. The study highlights the effectiveness of combining facial biometrics with environmental context using attention-enhanced recurrent networks. However, the authors note limitations regarding the small size and class imbalance of the Brain4Cars dataset. They suggest future research should focus on developing larger, more balanced datasets and implementing federated learning to address privacy concerns associated with using driver facial data.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | Crossref | — | — | 1 | 2026-06-19 |
| archive | success | canonical_url | — | — | 1 | 2026-06-25 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-19 |
| chunk | success | chunk | — | — | 1 | 2026-06-19 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-19 |
| promote | success | — | — | — | 1 | 2026-06-19 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-19 |
| verify | success | — | — | — | 1 | 2026-06-26 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; 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