Research map / Video / Video segmentation and tracking
DriveGazen: Event-Based Driving Status Recognition Using Conventional Camera
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DriveGazen, which converts conventional eye video frames into synthetic DVS event frames (via v2e) and classifies driving status with a guide-attention spiking network (ADSN), releasing the DriveGaze dataset.
Problem
Recognizing driver status (e.g., distraction, fatigue, anger) from eye observations under varying lighting, where conventional intensity frames lack temporal information.
Approach
Generate synthetic DVS event frames from conventional video with v2e, then process them with a spiking network whose guide attention module injects spatial texture cues from the first and last intensity frames into the event temporal features.
Key results
- On DriveGaze, Ours (E8-S7) reaches 92.4% WAR and 92.4% UAR, versus 91.3% for the second-best SEEN (E8-S7).
- On the SEE dataset, Ours (E4-S3) reaches 84.5% WAR/UAR, about 1% above SEEN (E4-S3) at 83.6%.
- Removing guide attention drops DriveGaze WAR from 91.2% to 65.4% for E4-S3 in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the first and last intensity frames rather than only the first frame (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 88.6 -> 91.2 | Helped |
| Use the first and last intensity frames rather than replacing the last frame with the second (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 89.2 -> 91.2 | Helped |
| Use the first and last intensity frames rather than all selected intensity frames (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 90.4 -> 91.2 | Helped |
| Add multi-head self-attention (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 89.5 -> 91.2 | Helped |
| Add spatial feature to spike feature fusion (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 90.5 -> 91.2 | Helped |
| Add the residual connection (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 84.0 -> 91.2 | Helped |
| Add guide attention (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 65.4 -> 91.2 | Helped |
| Use an SNN instead of a CNN temporal extractor (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 60.7 -> 91.2 | Helped |
| Use an SNN instead of an LSTM temporal extractor (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 60.8 -> 91.2 | Helped |
| Average membrane potentials rather than use only the last potential (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 88.5 -> 91.2 | Helped |
| Average membrane potentials rather than use only the last event frame's spikes (ablation) | DriveGazen · DriveGaze E4-S3 | WAR: 68.9 -> 91.2 | Helped |
Surprising findings
- Removing the guide attention module causes the largest ablation drop (to 65.4% WAR), far more than other components.
- Eyemotion slightly outperforms DriveGazen on low-light conditions in SEE, which the authors attribute to ImageNet pre-training.
Limitations
- Relies on events synthesized from intensity frames, so performance may degrade when intensity variations are minimal.
- Evaluated only on eye-region data from one wearable device and two datasets.
Experimental setup
Datasets: DriveGaze SEE
Benchmarks: DriveGaze SEE
Metrics: WAR UAR FLOPS inference time (ms)
Baselines: Eyemotion EMO SEEN ResNet18 + LSTM ResNet50 + GRU 3D ResNet18 R(2+1)D Former DFER
Builds on: v2e STBP LIF neuron model SEEN Eyemotion
Closest papers by meaning
- Autonomous Driving with Spiking Neural Networks (NeurIPS 2024 · similarity 0.84)
- On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance (NeurIPS 2024 · similarity 0.84)
- When Every Millisecond Counts: Real-Time Anomaly Detection via the Multimodal Asynchronous Hybrid Network (ICML 2025 · similarity 0.84)
- SalM²: An Extremely Lightweight Saliency Mamba Model for Real-Time Cognitive Awareness of Driver Attention (AAAI 2025 · similarity 0.83)
- DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.83)
- PhysDrive: A Multimodal Remote Physiological Measurement Dataset for In-vehicle Driver Monitoring (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.