Table 7.
Literature result (continued, Part 2 of 5): summary table of papers reporting on objective, analysis methods, and obtained results.
| Author(s) | Objective | Analysis Methods | Results |
|---|---|---|---|
| Dua et al., 2019a [49] | Detect and assess driver attention using the front camera of a windscreen-mounted smartphone | Neuronal networks, CNNs, and GRUs | The driver’s attention rating had an overall agreement of 0.87 with the ratings of 5 human annotators |
| Dua et al., 2019b [49] | Identify driver distraction based on facial characteristics (head position, eye gaze, eye closure, and yawning) | CNN (generic features) and GRU or (CNN + GRU) | The automatically generated rating has an overall agreement of 88% with the ratings provided by 5 human annotators; the attention-based model outperforms the AUTORATE model by 10% accuracy on the extended dataset |
| Eraqi et al., 2019 [50] | Detect 10 different types of driver distraction (including talking to passengers, phone calls, and texting) | Deep learning; ensemble of convolutional neural networks | New public dataset, detection with 90% accuracy |
| Gelmini et al., 2020 [46] | Driving style risk assessment based on speeding, longitudinal acceleration, lateral acceleration, and smartphone use while driving | Thresholds used for profiling drivers and detecting smartphone usage | Median phone usage, no accuracy indicators used |
| He et al., 2014 [80] | Present a seat-level location of smartphones in a vehicle to identify who is sitting where | Signal processing: reference frame transformation, event detection, left/right identification, front/back identification | Position accuracy between 70% and 90% (best case) |
| Hong et al., 2014 [81] | Detect a person’s driving behaviour via an Android-based in-vehicle sensor platform | Machine learning approach (Naïve Bayes classifier) | Average model accuracy with all three sensors was 90.5%, and 66.7% with the smartphone only |
| Janveja et al., 2020 [52] | Introduce a smartphone-based system to detect driver fatigue and distraction (mirror scanning behaviour) in low-light conditions | For distraction detection, statistics are calculated if the driver is scanning their mirrors at least once every 10 s continuously during the drive | NIR LED setup: 93.8% accuracy in detecting driver distraction |
| Jiao et al., 2021 [82] | Recognise actions of distracted drivers | Hybrid deep learning model, OpenPose, K-means, LSTM | Accuracy depending on processing step (up to 92%) |
| Johnson et al., 2011 [83] | Detect and classify driving events, such as left/right manoeuvres, turns, lane changes, device removal, and excessive speed and braking | Manoeuvre classification with the DTW algorithm | U-turn correctly identified 23% of the time (using accelerometer), 46% of the time (using gyroscope), 77% of the time (combined sensors), 97% of aggressive events correctly identified |
| Kapoor et al., 2020 [48] | Provide a real-time driver distraction detection system that detects distracting tasks in driver images | Convolutional neural networks (CNNs) | Accuracy for 4 classes (e.g., calling or texting on a cell phone) reaches 98–100% when fine-tuned with datasets such as the State Farm Distracted Driver Dataset |
| Kashevnik et al., 2021 [5] | Provide an audio-visual speech recognition corpus for use in speech recognition for driver monitoring systems | Corpus creation, development of smartphone app | Corpus (audio-visual speech database with list of phrases in Russian language, 20 participants) |
| Khurana and Goel, 2020 [84] | Detect smartphone use by drivers using in-device cameras | Random forest classifiers (machine learning models) for 2 scenarios: a) docked, b) in-hand | Approximately 90% accuracy in distinguishing between driver and passenger. Cannot collect data for phones in handheld position |
| Koukoumidis et al., 2011 [85] | Detect traffic lights using the smartphone camera and predict their timing | Machine learning (Support Vector Regression) | Accuracy of traffic signal detection (87.6% and 92.2%) and schedule prediction (0.66 s, for pre-timed traffic signals; 2.45 s for traffic-adaptive traffic signals) |