Generalized model for driver activity recognition in automated vehicles using pressure sensor array

Nobari, Khazar Dargahi; Bertram, Torsten · 2022 · Crossref

DOI: 10.54941/ahfe1002733

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 accurately recognizing driver activities in automated vehicles to facilitate efficient human-machine cooperation and ensure safety during takeover situations. Existing methods often rely on intrusive sensors or cameras that are sensitive to lighting, occlusion, and spatial constraints within the vehicle cabin. To overcome these limitations, the authors propose a generalized, subject-independent activity recognition model using unobtrusive seat pressure sensors. The goal is to detect both driving-related tasks and non-driving-related tasks (NDRTs) with high accuracy, thereby assessing driver readiness and engagement without requiring wearable devices or complex visual monitoring systems. The study was conducted using a static driving simulator equipped with two BodiTrak2 Pro pressure sensor mats (32 × 32 sensors each) placed on the seat and backrest. Data were collected from ten participants: eight selected via a fractional factorial design based on height, BMI, age, and gender for training, and two randomly selected for testing. Participants performed 20 distinct activities, including driving maneuvers (accelerating, braking, steering) and NDRTs such as eating, using mobile devices, reading, and relaxing, alongside a "no action" class. The pressure data, sampled at 16.7 Hz, were segmented into 3-second sliding windows. Three recurrent neural network architectures—bidirectional LSTM, stacked bidirectional LSTM, and CNN-LSTM—were trained to classify these 21 activity classes, leveraging their ability to process temporal sequential data. The results demonstrate that the proposed models achieved high accuracy despite using only pressure data from a limited number of training subjects. The bidirectional LSTM, stacked bidirectional LSTM, and CNN-LSTM models achieved accuracies of 90.2%, 91.3%, and 90.8%, respectively. These performance levels are comparable to or exceed state-of-the-art camera-based and WiFi-based approaches, which typically report accuracies between 81.6% and 94.0% but often require more intrusive setups or larger datasets. The study confirms that pressure distribution patterns provide sufficient information for robust activity recognition. The significance of this work lies in its validation of seat pressure sensors as a viable, non-intrusive alternative to camera-based monitoring for driver state assessment. By achieving high accuracy with a generalized model trained on a small, diverse dataset, the research supports the practical implementation of such systems in real-world vehicles. The findings suggest that integrating seat pressure data can enhance automated vehicle safety by reliably detecting driver engagement and readiness for takeover, while future work should focus on expanding the participant pool and fusing pressure data with other sensor sources to further improve classification performance.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-09; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness
• distraction detection algorithms

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).

• Methodological Resource: tool software
• Theoretical Contribution: computational model, conceptual framework