Source-linked AI summary
Real-time Distracted Driver Posture Classification
Yehya Abouelnaga, Hesham M. Eraqi, Mohamed N. Moustafa
TL;DR
Distracted-driver posture estimation requires recognizing safety-relevant activities from driver imagery. The paper introduces a new dataset and a genetically weighted CNN ensemble using raw, face, and hand views, achieving 95.98% accuracy while also examining posture confusions and real-time simplification.
Problem
Distracted driving causes crashes and injuries, motivating improved posture estimation for identifying driver distraction.
Method
The paper creates a distracted-driver dataset and combines AlexNet and InceptionV3 models trained on raw, face, hands, and face+hands images with genetically learned ensemble weights.
Results
95.98% classification accuracy was achieved by the genetically weighted ensemble on the distracted-driver posture task.
Takeaways & Limitations
Face and hands detection improved ensemble classification accuracy, while a simpler AlexNet model operated in realtime with satisfactory accuracy.
Takeaways & Limitations
The method remains constrained by imperfect face and hand detection and by missing temporal context for safe-driving confusions.
Abstract
from arXiv · showhide
In this paper, we present a new dataset for "distracted driver" posture estimation. In addition, we propose a novel system that achieves 95.98% driving posture estimation classification accuracy. The system consists of a genetically-weighted ensemble of Convolutional Neural Networks (CNNs). We show that a weighted ensemble of classifiers using a genetic algorithm yields in better classification confidence. We also study the effect of different visual elements (i.e. hands and face) in distraction detection and classification by means of face and hand localizations. Finally, we present a thinned version of our ensemble that could achieve a 94.29% classification accuracy and operate in a realtime environment.
1 Introduction
Distracted driving is a serious safety problem associated with crashes and injuries, motivating systems that estimate drivers’ distracted postures in real time.
- 3,477 people were killed and 391,000 were injured in 2015 motor-vehicle crashes involving distracted drivers.The NHTSA identifies activities such as phone use, eating, drinking, passenger conversation, and stereo operation as distractions.
- The paper presents a realtime distracted-driver pose-estimation system and evaluates it on a challenging distracted-driver dataset.
2 Literature Review
Prior distracted-driving research spans phone detection, indoor RGB-D studies, regional visual classifiers, public datasets, and handcrafted-feature classifiers.
- Research in distracted-driver detection has been organized into four groups, including cell-phone studies, UCSD datasets, Southeast University work, and StateFarm’s Kaggle competition.
- Earlier systems used SVMs, AdaBoost with Hidden Markov Models, and region-based classifier fusion for detecting phone use or driver activities.
- Zhao et al. evaluated Random Forests, k-nearest neighbors, and multilayer perceptrons on a side-view dataset with several driving activities.
- StateFarm introduced ten posture classes in the first publicly available posture-classification dataset, while this work created a similar dataset because access was restricted.
3 Dataset Design
The AUC Distracted Driver dataset was created to address limitations in existing datasets and contains examples of ten driving postures collected across participants and vehicles.
- The dataset represents ten postures, including drinking, radio adjustment, safe driving, grooming, reaching behind, passenger conversation, phone talking, and texting.
- The AUC dataset was needed because StateFarm’s data were restricted to competition use and the SEU dataset was unavailable and contained only four postures.
- Videos were recorded with a roof-mounted ASUS ZenPhone camera, converted into 1080 × 1920 images, and labeled using a custom annotation tool.
- 31 participants from seven countries contributed videos recorded in four different cars.
4 Proposed Method
The proposed method combines CNNs trained on multiple image sources and aggregates their class probabilities with genetically learned weights.
- 4 Proposed Method: The system trains AlexNet and InceptionV3 on raw, face, hands, and face+hands images, producing eight models whose outputs form the final class distribution.
- 4 Proposed Method: InceptionV3 uses transfer learning from a pretrained ImageNet model, while AlexNet is trained from scratch and outputs ten posture classes.
- 4.3 Weighted Ensemble of Classifiers using Genetic Algorithm: Each classifier produces a ten-probability softmax vector, and the ensemble combines the vectors using weighted aggregation across N = 8 classifiers.
- 4.3 Weighted Ensemble of Classifiers using Genetic Algorithm: A genetic algorithm estimates classifier weights because the method does not assume that all classifiers contribute equally.
- 4.3 Weighted Ensemble of Classifiers using Genetic Algorithm: The genetic algorithm uses classifier weights as chromosome genes and evaluates Negative Log Likelihood on a 50% random population sample to reduce overfitting.
5 Experiments
Experiments evaluate posture classifiers on held-out data, comparing image inputs and analyzing ensemble errors. Static images particularly confuse safe driving with distracted activities because temporal context is unavailable.
- 75% of the dataset was used for training and 25% for held-out testing.
- AlexNet and InceptionV3 achieved their best accuracies when trained on original images, while hands contributed more to posture recognition than faces.Face + Hands images produced slightly lower accuracy than hands images, yet remained higher than the truncated comparison in the passage.
- The most confusing posture was “safe driving,” because static images lack temporal context and can hide concurrent distraction.
- “Text Left” and “Text Right” were mainly confused with their corresponding talking postures, while “Adjust Radio” was mainly confused with safe driving.
- “Drink” and “Talk to Passenger” were comparatively well classified, with 98% and 97.67% of images correctly classified, respectively.
6 Conclusion
The paper presents a vision-based distracted-driving posture system and a challenging dataset, with a genetically weighted CNN ensemble achieving 95.98% accuracy. It also identifies realtime overhead and missing temporal context as areas for improvement.
- 95.98% accuracy was achieved by the best model, a genetically weighted ensemble of convolutional neural networks.
- A simpler AlexNet-only model operated in realtime while maintaining satisfactory classification accuracy.
- Face and hands detection improved ensemble classification accuracy, but their realtime performance overhead was much higher than their contribution.
- Future work proposes better face and hands detection and recurrent models to incorporate temporal information and address safe-driving confusion.