YOLO-APD: Adaptive Pedestrian Detector

• The paper enhances YOLOv8 with SimAM attention, C3Ghost, SimSPPF, Mish activation, and an IGD-inspired fusion structure for robust pedestrian detection on curved roads.
• It integrates a steering-aware dynamic ROI mechanism that focuses on safety-critical regions, achieving 97.0% [email protected] and 77.7% [email protected]:0.95 at 100 FPS on a CARLA benchmark.
• The approach addresses key challenges in monocular RGB detection such as occlusion, scale variation, and limited field-of-view in geometrically complex driving environments.

Searching arXiv for the cited paper and closely related YOLO adaptive detection work. YOLO-APD is a YOLOv8-derived pedestrian detector specialized for autonomous driving in visually difficult, geometrically complex road environments, especially Type-S curved roads, where restricted field of view, clutter, occlusion, scale variation, and rapid viewpoint changes degrade conventional monocular RGB detection. The method enhances YOLOv8 through a specific combination of SimAM attention, C3Ghost modules, a SimSPPF replacement for SPPF, Mish activation, and an Intelligent Gather & Distribute (IGD)-inspired fusion structure, while also proposing a steering-aware dynamic region-of-interest mechanism for curve-aligned perception. On a custom CARLA benchmark, the paper reports 97.0% [email protected], 77.7% mAP@0.5:0.95, pedestrian recall exceeding 96%, and 100 FPS, positioning the model as a real-time, RGB-only detector for challenging road geometries (Joctum et al., 7 Jul 2025).

1. Problem domain and design rationale

YOLO-APD addresses pedestrian detection for autonomous driving on geometrically complex road environments, with particular emphasis on Type-S curved roads. In the paper’s formulation, these roads create repeated visibility constraints, frequent partial occlusions, and reduced reaction time because pedestrians may emerge abruptly from behind bends, roadside clutter, or vehicles. The method is explicitly motivated by the need for a cost-effective monocular RGB-camera solution, rather than a heavier stack combining LiDAR, RADAR, and multiple cameras (Joctum et al., 7 Jul 2025).

The target failure modes are stated clearly. Standard RGB-only detection is described as difficult under reduced visibility on curves and blind corners, occlusion and background clutter, scale variation, adverse weather and low illumination, and rapid viewpoint changes caused by vehicle motion. The paper further argues that even strong single-stage baselines such as YOLOv8 may degrade in these conditions, especially for small-scale pedestrians, multi-scale feature degradation, insufficient feature fusion across scales, and restricted field of view and dynamic visibility shifts (Joctum et al., 7 Jul 2025).

The resulting design goals are correspondingly narrow and operational. YOLO-APD is constructed to improve robustness in cluttered and curved scenes, strengthen small-object and multi-scale pedestrian detection, improve cross-scale fusion, preserve real-time speed, and maintain deployment on RGB input alone. This focus distinguishes it from more general-purpose YOLO modifications: the architecture is organized around the safety-critical constraint that false negatives on pedestrians are unacceptable. A plausible implication is that recall is treated not merely as a performance metric but as a deployment-oriented objective.

2. Architectural composition

The baseline is YOLOv8, and the paper later states that YOLO-APD outperforms its direct baseline YOLOv8m. The network pipeline resizes input images to 640×640 and modifies the backbone, neck, and activation design while remaining within the YOLO detection paradigm (Joctum et al., 7 Jul 2025).

In the backbone, YOLO-APD introduces SimAM attention, replaces heavier bottleneck/C3-style modules with C3Ghost, substitutes the standard SPPF with SimSPPF, and uses Mish activation instead of more conventional activations. In the neck, the default YOLOv8 FPN/PAN-style fusion path is replaced by an IGD-inspired neck/head fusion structure with components shown in the paper as IFM, Inject / InjectMulti, SimFusion 3-in, SimFusion 4-in, AdvPoolFusion, TopBasicLayer, and bilinear upsampling and pooling-based aggregation. The head remains a YOLO-style detection head, but it now operates on features produced by this richer fusion pathway (Joctum et al., 7 Jul 2025).

Several of these substitutions are motivated by complementary constraints. SimAM is described as a parameter-free attention module that assigns importance weights to neurons across spatial and channel dimensions without adding learnable parameters. C3Ghost combines Cross-Stage Partial structure with Ghost convolutions so that a smaller set of intrinsic feature maps is produced by normal convolution and additional “ghost” maps are generated by cheaper linear operations, reducing FLOPs and parameter count. SimSPPF is intended to recover stronger multi-scale contextual information than standard SPPF, while Mish is used to improve optimization stability and generalization. The IGD-inspired neck is meant to overcome the limitations of adjacent-scale-only information passing in standard FPN/PAN fusion (Joctum et al., 7 Jul 2025).

The paper also proposes a dynamic region-of-interest (DROI) mechanism guided by steering angle and vehicle speed. This mechanism is presented as a system-level extension rather than a fully integrated component of the tested detector. Its role is to define a trajectory-relevant ROI ahead of the vehicle, especially in curves, so that perception can be focused on the most safety-critical region. A common misconception is to read this DROI as part of the experimentally validated architecture; the paper explicitly treats it as a conceptual extension and future deployment direction rather than a fully implemented benchmarked module (Joctum et al., 7 Jul 2025).

3. Principal modules and mathematical formulation

The paper provides explicit mathematical detail for several modules. For SimAM, the attention mechanism is derived from an energy function over a feature map $X \in \mathbb{R}^{C \times H \times W}$, where lower energy corresponds to greater separability of a target neuron from its surroundings. The attention-weighted feature is written as

$X' = \sigma\!\left(\frac{1}{E}\right)\odot X$

with $\sigma(\cdot)$ denoting the sigmoid and $\odot$ elementwise multiplication (Joctum et al., 7 Jul 2025). The intended effect is to highlight informative neurons while suppressing background responses in cluttered road scenes.

For Mish, the paper gives

$f(x) = x \tanh(\operatorname{softplus}(x))$

and equivalently

$f(x) = x\tanh(\ln(1+e^x)).$

This smooth, non-monotonic activation is presented as improving optimization and information propagation relative to ReLU-like alternatives (Joctum et al., 7 Jul 2025).

For SimSPPF, the paper does not provide a single closed-form equation, but it specifies the computation sequence: a first SimConv, followed by three serial $5 \times 5$ max-pooling operations with stride $1$ and padding $2$, concatenation of the original transformed feature with the three pooled outputs, and a second SimConv. The module uses SimConv = Conv + BN + Mish. This design is intended to create progressively larger effective receptive fields and then fuse the resulting multi-scale context (Joctum et al., 7 Jul 2025).

The detection loss is given as

$L_{\text{total} = \lambda_1 L_{\text{cls} + \lambda_2 L_{\text{box} + \lambda_3 L_{\text{DFL}$

where classification uses BCEWithLogitsLoss, box regression uses CIoU loss, and localization refinement includes Distribution Focal Loss (DFL). The paper writes CIoU as

$X' = \sigma\!\left(\frac{1}{E}\right)\odot X$0

with $X' = \sigma\!\left(\frac{1}{E}\right)\odot X$1, and gives DFL as

$X' = \sigma\!\left(\frac{1}{E}\right)\odot X$2

These equations place YOLO-APD within the standard modern YOLOv8 localization regime rather than introducing a new training objective from scratch (Joctum et al., 7 Jul 2025).

The steering-aware ROI width is expressed as

$X' = \sigma\!\left(\frac{1}{E}\right)\odot X$3

where $X' = \sigma\!\left(\frac{1}{E}\right)\odot X$4 is the base ROI width, $X' = \sigma\!\left(\frac{1}{E}\right)\odot X$5 is steering angle, and $X' = \sigma\!\left(\frac{1}{E}\right)\odot X$6 is vehicle speed. The paper describes piecewise behavior in which the critical region remains near $X' = \sigma\!\left(\frac{1}{E}\right)\odot X$7 under straight driving, expands laterally under moderate curves, and extends more strongly along the predicted path during sharp turns (Joctum et al., 7 Jul 2025).

A notable omission is that exact mathematical equations are not specified for C3Ghost generation or for the detailed fusion operations inside IGD, including SimFusion 3-in, SimFusion 4-in, Inject, AdvPoolFusion, and IFM. The architecture is therefore better specified at the systems level than at the algebraic level.

4. Experimental methodology and quantitative performance

The principal benchmark is a synthetic CARLA v0.9.15 dataset generated on Unreal Engine 4 using Town02–Town05 maps. The scenarios span daylight, twilight, and night, weather conditions including clear, rainy, and foggy, and varying densities of dynamic vehicles and pedestrians. Images are captured from a forward-facing RGB camera mounted on an ego vehicle in autopilot mode at 640×480 resolution. The dataset contains 2015 images, split into 1753 training, 183 validation, and 79 test images, with 10 annotated classes: person, car, truck, bike, motorbike, traffic light green, traffic light red, traffic light orange, traffic sign 30, and traffic sign 90. The paper also notes inconsistencies in later reporting, where some figures mention different class names or what appears to be 8 or 10 categories (Joctum et al., 7 Jul 2025).

Training uses AdamW, learning rate 0.01, batch size 16, weight decay 0.0005, momentum 0.937, COCO pre-trained weights, and augmentation limited in the reported text to mosaic and horizontal flips. Hardware and software details are also given: Windows, Intel Core i5-11400H, 16 GB RAM, Tesla P100-PCIE, Python 3.8.0, PyTorch 1.8.0, Visual Studio Code, Unreal Engine 4.27, and CARLA 0.9.15 (Joctum et al., 7 Jul 2025).

The headline CARLA results are summarized below.

Model	[email protected]	[email protected]:0.95	FPS
YOLOv8	93.1	71.5	~120
YOLO-APD	97.0	77.7	100

The paper additionally reports mF1 = 0.944, FLOPs = 76.5 G, and Parameters = 24.16 M for YOLO-APD, versus 67.7 GFLOPs and 20.04 M params for YOLOv8. Relative to YOLOv8, YOLO-APD improves [email protected] by about +3.9 points and [email protected]:0.95 by about +6.2 points, while remaining real-time at 100 FPS (Joctum et al., 7 Jul 2025).

The paper also compares against SSD, Faster R-CNN, YOLOv5, and YOLOv7. SSD is reported at 86.4 [email protected] and 60.5 mAP; Faster R-CNN at 90.1 and 62.5 with approximately 5 FPS; YOLOv5 at 90.8 and 66.1; and YOLOv7 at 93.2 and 67.7. These comparisons underpin the claim that YOLO-APD achieves the strongest overall balance of accuracy and practical throughput on the custom benchmark (Joctum et al., 7 Jul 2025).

The confusion-matrix discussion reports especially strong results for the person class, with normalized precision and recall of 1.00 on the test set. The paper nonetheless emphasizes the more conservative headline that pedestrian recall exceeds 96%, which is more appropriate given the small test split and class-specific composition (Joctum et al., 7 Jul 2025).

5. Ablation, transfer behavior, and limitations

The ablation study proceeds cumulatively from baseline YOLOv8. The reported progression is Exp1 baseline YOLOv8 with AP = 0.93082 and mAP = 0.71453; Exp2 adding SimSPPF to reach 0.95188 / 0.72907; Exp3 adding SimAM to reach 0.95341 / 0.73152; Exp4 adding IGD to reach 0.95244 / 0.74564; and Exp5 adding Mish to reach 0.97053 / 0.77444. The paper interprets these results as showing that SimSPPF provides the strongest early gain, SimAM adds a modest but positive refinement, IGD-inspired fusion improves stricter localization quality more than loose-threshold AP, and Mish delivers the final major performance jump (Joctum et al., 7 Jul 2025).

This ablation also clarifies an important technical limitation. Although C3Ghost is listed as part of the architecture and discussed qualitatively as contributing to efficiency, the paper does not provide a separate ablation row for C3Ghost alone. Consequently, its isolated numerical contribution is not established in the same way as SimSPPF, SimAM, IGD, or Mish. A common misconception would be to assume that every listed module has equally direct ablation support; the paper does not show that (Joctum et al., 7 Jul 2025).

For real-world transfer, the authors evaluate on KITTI Object Detection, focusing on Person and Car, with images processed at 640×640. The split is 6732 training and 749 validation images, but the paper states that YOLO-APD is trained solely on CARLA and evaluated on KITTI for transferability analysis. Reported KITTI results are Person: AP = 0.965, F1 = 0.817 and Car: AP = 0.976, F1 = 0.954, compared with CARLA values of Person: AP = 0.981, F1 = 0.9904 and Car: AP = 0.934, F1 = 0.9249 (Joctum et al., 7 Jul 2025).

The interpretation is explicit: synthetic-to-real transfer is incomplete, especially for pedestrians, with a small AP drop but a substantial F1 drop on KITTI. The paper attributes this to domain shift and class imbalance, especially fewer pedestrian instances in KITTI. It therefore concludes that robust domain adaptation is still needed, mentioning fine-tuning on target-domain data, domain randomization, domain-invariant feature learning, and GAN-based synthetic-to-real translation as future remedies (Joctum et al., 7 Jul 2025).

Several limitations are acknowledged or directly evident. The main benchmark is synthetic; the DROI mechanism is conceptual rather than fully integrated experimentally; implementation details for IGD are incomplete; class-label reporting contains minor inconsistencies; background false positives remain for bike and vehicle; and YOLO-APD has higher parameters, computation, and inference time than the base YOLOv8, even if the increase is modest relative to the reported accuracy gain (Joctum et al., 7 Jul 2025).

6. Position within adaptive YOLO research

Within the broader literature, YOLO-APD is best understood as an application-targeted architectural enhancement of YOLOv8, not as a general MoE router, a UDA framework, or a domain-pretraining pipeline. The paper itself states that the novelty lies primarily in the specific integration of SimAM, C3Ghost, Mish, SimSPPF, and an IGD-inspired fusion structure for pedestrian detection on curved roads, rather than in inventing all of these submodules independently (Joctum et al., 7 Jul 2025).

This places it in a distinct position relative to other adaptive YOLO variants on arXiv. "YOLO Meets Mixture-of-Experts: Adaptive Expert Routing for Robust Object Detection" builds a routed multi-expert detector from multiple YOLOv9-T experts, with two experts, three routers total, feature-conditioned routing, output-level fusion, and a load-balancing term, reporting gains on mixed COCO + VisDrone training; unlike YOLO-APD, it is an MoE system with soft mixture inference and higher detector-level redundancy rather than a single enhanced backbone-neck-head design (Meiraz et al., 17 Nov 2025). "CLDA-YOLO: Visual Contrastive Learning Based Domain Adaptive YOLO Detector" addresses unsupervised domain adaptive object detection through a teacher-student cooperative system, uncertainty-aware pseudo-label distillation, dynamic augmentation, and multi-stage contrastive alignment, explicitly aiming to preserve inference speed by using adaptation modules only during training; this is directly relevant to YOLO-APD’s observed CARLA-to-KITTI gap, because it targets precisely the kind of domain shift that YOLO-APD identifies as unresolved (Qiu et al., 2024). "PK-YOLO: Pretrained Knowledge Guided YOLO for Brain Tumor Detection in Multiplanar MRI Slices" instead adapts YOLOv9-E to medical imaging through Spark-pretrained RepViT and Focaler-IoU, showing that in-domain self-supervised pretraining and small-object-aware regression can materially alter YOLO behavior in a specialized domain; unlike YOLO-APD, its adaptation strategy is domain-specific pretraining rather than road-scene feature-fusion redesign (Kang et al., 2024).

Taken together, these comparisons clarify the scope of YOLO-APD. It is a detector specialized for robust pedestrian detection on complex road geometries, with its strongest contributions in multi-scale feature capture, salient-feature emphasis, cross-scale fusion, and real-time RGB-only deployment. It does not, in its current form, solve expert routing, synthetic-to-real domain adaptation, or training-time domain knowledge injection as comprehensively as those other lines of work. This suggests that future extensions of YOLO-APD could plausibly combine its curved-road-specific architecture with explicit domain adaptation or adaptive routing mechanisms, but such combinations are not part of the reported method.