YOLO11-CR: Real-time Fatigue Driving Detector

Updated 9 July 2026

YOLO11-CR is a lightweight, YOLO11-derived detector tailored for real-time fatigue driving detection using in-vehicle RGB video.
The model integrates a Convolution-and-Attention Fusion Module (CAFM) and Rectangular Calibration Module (RCM) to enhance detection of occluded, small, and elongated objects.
Quantitative evaluations show high precision and efficient throughput, with ablation studies confirming CAFM’s occlusion sensitivity and RCM’s better localization.

to=arxiv_search.search 在天天中彩票天天爱彩票怎么_code {"query":"(Jin et al., 16 Aug 2025) YOLO11-CR Lightweight Convolution-and-Attention Framework Accurate Fatigue Driving Detection; (Kotthapalli et al., 4 Aug 2025) YOLOv1 to YOLOv11 survey YOLO11-CR", "max_results": 5} to=arxiv_search.search 仲博ിറ്റেম_code {"query":"(Jin et al., 16 Aug 2025)", "max_results": 3} YOLO11-CR is a lightweight YOLO11-derived object detector for vision-based fatigue driving detection that builds on YOLO11n and targets real-time, non-intrusive monitoring in intelligent transportation systems. In its canonical usage, the name refers to the model introduced for detecting three primary targets in in-vehicle RGB video—normal face, profile face, and mobile phone—through two architectural additions: the Convolution-and-Attention Fusion Module (CAFM) and the Rectangular Calibration Module (RCM). The design addresses small or occluded objects, side-view faces with elongated contours, and limited multi-scale context modeling while preserving a lightweight compute profile suitable for real-time inference (Jin et al., 16 Aug 2025).

1. Definition and problem setting

YOLO11-CR is defined as a lightweight and efficient object detection model tailored for real-time fatigue detection. Its stated application domain is fatigue-related driver behavior analysis from RGB in-vehicle video, with emphasis on non-intrusive deployment and robustness to small or occluded objects such as mobile phones, as well as profile faces under diverse lighting and pose conditions (Jin et al., 16 Aug 2025).

The model is motivated by three recurring constraints in vision-based fatigue monitoring. First, standard CNN-based detectors can struggle with small or partially occluded targets, including phones partially hidden by hands or steering wheels. Second, purely convolutional feature extractors provide limited global context, which weakens performance when fatigue cues are spatially distributed or partially occluded. Third, heavy attention or multi-branch designs can compromise latency and memory efficiency, making embedded deployment difficult. YOLO11-CR addresses these issues by modifying YOLO11n rather than replacing the baseline detection pipeline outright (Jin et al., 16 Aug 2025).

Within the detector’s target taxonomy, the three operational classes are normal face, profile face, and mobile phone. The broader DSM dataset used in the work contains annotations for “normal driving,” “eye closure,” “yawning,” “head dropping,” “phone usage,” and “looking sideways,” but the detection formulation adopted for YOLO11-CR focuses specifically on the three object classes above. This design suggests an object-centric formulation of fatigue monitoring in which fatigue-relevant states are proxied through face orientation and phone-use cues rather than direct sequence-level behavior classification (Jin et al., 16 Aug 2025).

2. Architectural composition

YOLO11-CR uses YOLO11n as its base detector and retains the standard encoder-decoder organization of backbone, neck, and head. The backbone produces multi-scale features at P3, P4, and P5, corresponding to strides 8, 16, and 32. The neck preserves upsampling, concatenation, CBS/C3k2 blocks, and SPPF, while the head remains a multi-scale, anchor-free detection head with decoupled classification and box regression branches (Jin et al., 16 Aug 2025).

The principal structural change is the insertion of CAFM into the late backbone and the insertion of RCM into the neck and detection heads. CAFM replaces the attention layer inside the C2PSA block of YOLO11n, forming a “C2PSA_CAFM” structure. RCM replaces conventional $3\times 3$ and $1\times 1$ convolutions in specific feature extraction and fusion stages, yielding “C3k2_RCM” blocks in the neck, and an additional RCM block is embedded in each detection scale at P3, P4, and P5 (Jin et al., 16 Aug 2025).

CAFM is a dual-branch module intended to fuse local convolutional detail with global self-attention context. Its local branch applies a $1\times 1$ convolution, channel shuffle, and a $3\times 3$ depthwise convolution:

$F_{\mathrm{conv}} = W_{3\times 3}^{dw}\!\big(\mathrm{CS}(W_{1\times 1}(Y))\big).$

Its global branch computes channel-wise scaled dot-product attention through learned $Q$ , $K$ , and $V$ projections:

$\mathrm{Attention}(Q,K,V) = V \cdot \mathrm{Softmax}\!\left(\frac{K^\top Q}{\alpha}\right),$

followed by residual projection,

$F_{\mathrm{attn}} = W_{1\times 1}\big(\mathrm{Attention}(Q,K,V)+Y\big),$

and final fusion,

$1\times 1$ 0

The module uses depthwise convolutions and $1\times 1$ 1 projections to limit parameter and FLOP growth (Jin et al., 16 Aug 2025).

RCM is designed for axis-aware spatial calibration, especially for elongated or rectangular targets such as profile faces and mobile phones. It first aggregates horizontal and vertical context through horizontal pooling and vertical pooling:

$1\times 1$ 2

It then applies strip convolutions for shape self-calibration,

$1\times 1$ 3

followed by depthwise local refinement and residual output:

$1\times 1$ 4

$1\times 1$ 5

The intended effect is to improve alignment for profile faces and small phones by explicitly encoding horizontal and vertical contextual structure (Jin et al., 16 Aug 2025).

At the system level, the model remains lightweight despite these additions. The reported complexity is 2.691M parameters and 3.28 GLOPs for YOLO11-CR, compared with 2.624M parameters and 3.31 GLOPs for YOLO11n, indicating a slight parameter increase but a slight FLOP decrease (Jin et al., 16 Aug 2025).

3. Dataset, training protocol, and evaluation methodology

YOLO11-CR is evaluated on the DSM dataset derived from DMD. The dataset contains more than 180,000 RGB images from in-vehicle video recorded in urban, highway, and night-driving conditions, with dashboard and ceiling cameras capturing frontal and profile views. The dataset annotations include bounding boxes, facial landmarks, and occlusion levels, and the full behavior label distribution is reported as 45,000 normal driving images, 36,000 eye closure, 30,000 yawning, 27,000 head dropping, 21,000 phone usage, and 21,000 looking sideways. The split is train/validation/test = 7:2:1 (Jin et al., 16 Aug 2025).

Training is performed with SGD, an initial learning rate of 0.001, cosine annealing scheduling, momentum 0.937, batch size 64, and 100 epochs. The augmentation policy reported in the paper consists of Mosaic and random horizontal flip. The software and hardware stack is Windows 11, Python 3.11.2, PyTorch 2.5.1 with CUDA 12.4, an AMD Ryzen 9 5950X CPU, an NVIDIA RTX 3090 GPU, and 64 GB DDR5 RAM (Jin et al., 16 Aug 2025).

The paper states that YOLO11-CR uses the default YOLO11 training and inference procedure and does not enumerate altered loss functions or non-maximum suppression variants. The reported evaluation metrics are Average Precision, mean Average Precision, Precision, and Recall, with the following formulas:

$1\times 1$ 6

$1\times 1$ 7

$1\times 1$ 8

$1\times 1$ 9

The protocol defines mAP@50 at IoU threshold 0.5 and mAP@50–95 as the average over IoU thresholds from 0.5 to 0.95 (Jin et al., 16 Aug 2025).

4. Quantitative performance and ablation findings

On the DSM test set, YOLO11-CR achieves Precision 87.17%, Recall 83.86%, mAP@50 88.09%, and mAP@50–95 55.93%. In the same evaluation, YOLOv8 reports Precision 0.740, Recall 0.796, mAP@50 0.769, and mAP@50–95 0.499; YOLOv10n reports 0.856, 0.761, 0.842, and 0.543; and YOLO11n reports 0.865, 0.762, 0.823, and 0.534. The corresponding efficiency figures are 3.157M parameters and 4.43 GLOPs for YOLOv8, 2.775M and 4.37 for YOLOv10n, 2.624M and 3.31 for YOLO11n, and 2.691M and 3.28 for YOLO11-CR (Jin et al., 16 Aug 2025).

Model	Accuracy	Efficiency
YOLOv8	P 0.740, R 0.796, mAP@50 0.769, mAP@50–95 0.499	3.157M params, 4.43 GLOPs, 184.0 FPS
YOLOv10n	P 0.856, R 0.761, mAP@50 0.842, mAP@50–95 0.543	2.775M params, 4.37 GLOPs, 111.1 FPS
YOLO11n	P 0.865, R 0.762, mAP@50 0.823, mAP@50–95 0.534	2.624M params, 3.31 GLOPs, 129.6 FPS
YOLO11-CR	P 0.872, R 0.839, mAP@50 0.881, mAP@50–95 0.559	2.691M params, 3.28 GLOPs, 106.4 FPS

Per-class results further localize the detector’s strengths and failure modes. For Normal Face, YOLO11-CR reports Precision 0.930, Recall 0.959, mAP@50 0.986, and mAP@50–95 0.821. For Profile Face, it reports 0.909, 0.754, 0.891, and 0.483. For Mobile Phone, it reports 0.758, 0.659, 0.716, and 0.366. These figures identify the mobile phone category as the most difficult class, with the lowest localization accuracy under stricter IoU thresholds (Jin et al., 16 Aug 2025).

The ablation study attributes distinct roles to the two added modules. Starting from YOLO11n as baseline, adding RCM yields Precision 0.8810, Recall 0.7748, mAP@50 0.8494, and mAP@50–95 0.5358; adding CAFM yields Precision 0.8466, Recall 0.7942, mAP@50 0.8416, and mAP@50–95 0.5363; combining both produces YOLO11-CR at 0.8717, 0.8386, 0.8809, and 0.5593. The paper interprets RCM as primarily improving localization and box alignment, while CAFM mainly improves sensitivity to challenging or occluded objects through stronger global context. Their combination yields the best aggregate performance (Jin et al., 16 Aug 2025).

Qualitative analysis in the same study indicates that YOLO11-CR maintains higher precision at high recall, improves correct classification rates for profile faces and mobile phones, and reduces failure on small or occluded objects relative to the baselines. Confusion-matrix analysis reports 96% correctness for profile face, 84% for normal face, and 85% for mobile phone, with the phone class showing the largest improvement over the comparison models (Jin et al., 16 Aug 2025).

5. Real-time deployment profile, limitations, and future directions

YOLO11-CR is explicitly presented as a real-time model. Its reported throughput is 106.4 FPS on an RTX 3090, and its 2.691M-parameter, 3.28-GLOP profile is described as remaining lightweight despite the added modules. The paper argues that the small FLOP delta relative to YOLO11n supports deployment on embedded GPUs and potentially mobile NPUs after additional optimization, although quantization and pruning are proposed only as future work rather than evaluated components of the reported system (Jin et al., 16 Aug 2025).

The principal deployment rationale is that CAFM and RCM improve the representation of small or occluded fatigue-related targets without introducing a large computational penalty. RCM’s axial pooling and strip convolutions are linked to better localization of rectangular or elongated objects, while CAFM’s local-global fusion mitigates the local-only biases of standard CNN extraction. This suggests a detector that is optimized not merely for class discrimination but for the spatial morphology of in-cabin fatigue cues (Jin et al., 16 Aug 2025).

Several limitations remain explicit. The system is frame-based, so temporal fatigue patterns such as blink dynamics or microsleep progression are not modeled. Mobile phones remain the hardest class, with mAP@50–95 of 36.6%. Background variability can still induce false positives at permissive thresholds. The authors identify temporal modeling, multi-modal fusion with thermal or physiological signals, and stronger domain generalization across cameras and environments as major future directions. They also note that more advanced post-processing and training modifications—such as label smoothing, EMA, adaptive NMS, or refined box losses—were not explored in the reported experiments (Jin et al., 16 Aug 2025).

Although YOLO11-CR has an exact title match in the fatigue-driving paper above, the suffix “CR” is not stable across the broader YOLO11 literature. A comprehensive YOLO survey states that “the paper does not explicitly use or define the term ‘YOLO11-CR’” and treats “CR” as a reasonable interpretation of “Classification–Regression,” mapping it to a decoupled-head YOLOv11 variant rather than to a distinct published model (Kotthapalli et al., 4 Aug 2025). A clinical ultrasound study likewise states that the exact term does not appear in the paper and informally maps it to a “clinical-ready” YOLO11s deployment for intraoperative brain tumor detection (Cepeda et al., 27 Jan 2025).

Outside those interpretive uses, the suffix has been repurposed for domain-specific meanings. In vehicle metadata extraction, it is mapped to a YOLO-v11 color-recognition configuration with multi-view inference (Al-Saddik et al., 25 Jul 2025). In cotton detection, it is treated as “YOLO11—Cotton Recognition,” with the customized detector COTONET described as a YOLO11-derived cotton-specific model (González et al., 12 Mar 2026). In resource-optimization work, it is associated with “compressed/compute-reduced” variants that prune YOLOv11 heads according to object-size profiles (Rasheed et al., 2024). In railroad foreign-object detection, the term is interpreted through “channel reconstruction and reduction,” particularly in connection with SC_Detect and LAMP pruning (Liu et al., 12 Oct 2025). A crack-detection study also uses “YOLO11-CR” as a descriptive label for a YOLOv11-based crack recognition model realized as YOLO-AMC (Tsai et al., 11 Jun 2026).

Taken together, these usages indicate that “YOLO11-CR” is polysemous in adjacent literature. In strict bibliographic terms, the only explicit model carrying the exact name is the fatigue-driving detector built from YOLO11n with CAFM and RCM (Jin et al., 16 Aug 2025). In neighboring papers, the same suffix functions as an interpretive shorthand rather than a standardized architecture label (Kotthapalli et al., 4 Aug 2025).