To address the joint requirements of illegal parking detection and license plate recognition under complex traffic scenarios and limited edge-device resources, this study proposes RKF-YOLO, a lightweight dual-task model based on improved YOLOv11n that integrates Rep-CSP structural optimization, knowledge-transfer-enhanced training (KTET), and Focal-CIoU loss. Compared with YOLOv11n, RKF-YOLO reduces parameters and FLOPs by 38.2% and 38.1%, respectively, while improving $\mathrm{mAP}@0.5$ and $\mathrm{mAP}@0.5:0.95$ by 0.6 and 1.1 percentage points for parking detection; for plate detection, Focal-CIoU improves $\mathrm{mAP}@0.5:0.95$ by 1.3 percentage points and contributes to a recognition accuracy of 95.7%. The unified framework uses a shared backbone and task-oriented detection heads to support vehicle-level illegal parking detection and license-plate-oriented localization. Rep-CSP enhances multi-scale feature representation, asymmetric channel reduction with feature compensation reduces redundant computation, and KTET improves convergence through optimizer and learning-rate migration. Deployment on RK3588 achieves 59.5 FPS for parking detection and 95.1% recognition accuracy, demonstrating real-time performance and practical applicability on resource-constrained edge devices.

Traditional indoor localization using received signal strength (RSS) fingerprinting commonly uses measured RSS values as location-related inputs. This paper investigates a paired light–dark representation of an RSS fingerprint image and a cross-modality multi-scale fusion model for small-scale indoor localization. Specifically, each RSS fingerprint image is inverted in pixel intensity, and the original and inverted images are retained as synchronized representations of the same RSS measurement rather than as independent physical modalities or separate augmented samples. The paired images are processed by two parallel convolutional neural network (CNN) paths combined with a cross-modality fusion (CMF) module, which extracts and fuses multi-scale feature representations. The fused features are fed into fully connected layers to predict the coordinate vector. Experimental comparisons with traditional and advanced methods on four public datasets show positioning errors comparable to the strongest advanced benchmark and lower errors than the considered traditional baselines. The proposed model uses approximately 30.8% fewer parameters than the advanced benchmark.

Bandwidth-limited aerial–ground sensing requires an explicit trade-off among payload size, decoded-map quality, and downstream route utility. This paper evaluates a strict RGB-D edge-perception interface in which the receiver accesses only a quantized C × 16 × 16 tensor and no encoder-side skip features. Under a fixed AeroScapes protocol, five-seed latent-8 training gives a test intersection over union (IoU) of 0.5617 ± 0.0585 and route utility of 0.0539 ± 0.0227; the strongest validation-selected checkpoint reaches an IoU of 0.6584 but is reported only as a checkpoint-specific result. Acontrolled five-seed width ablation identifies a Pareto set: latent-8 is the lowest-rate operating point at 16.448 kbit, latent-16 has the smallest IoU standard deviation, and latent-32 gives the highest mean IoU (0.5788) and route utility (0.0921), with no significant pairwise differences between widths. A paired threeseed stabilization test likewise finds no significant IoU improvement from depth dropout, a soft topology loss, or their combination (p ≥ 0.6060); the combined configuration raises mean IoU to 0.5700 but increases dispersion. Relative to a practical reference combining JPEG (quality 75) RGB and 8-bit PNG depth, latent-8 reduces the mean payload by a factor of 4.32. The digital-link study extends the additive-noise analysis to fading, intersymbol interference, timevariation, packet and burst errors, near–far interference, and corrupted range metadata. Conventional short codes, source-aware unequal protection, and a 3GPP NR LDPC implementation are evaluated with framing, automatic repeat request, mediumaccess overhead, and transmission delay. Cross-domain evaluation establishes an important limitation: zero-shot AeroScapes-to-UAVid transfer is weak, whereas sequence-disjoint tenseed UAVid training improves the eight-class mean IoU from 0.2417 ± 0.0221 for RGB-only to 0.2602 ± 0.0129 for strict RGB-D (p = 0.0161). Lightweight depth, visibility, missingdepth, weighted-topology, and coarse-to-local refinement experiments further delimit deployment. Together, these experiments provide an auditable cross-layer evaluation linking source representation, channel reliability, protocol cost, spatial error, and receivergrid connectivity without equating one favorable checkpoint with general superiority.