DensePed-Lite: Quality-Aware Adaptive Detection for Dense Pedestrians under Occlusion

• Lightweight detectors lack adaptation to occlusion-induced ambiguity, which manifests as unreliable confidence, fragmented spatial features, and limited scanning paths.Source: §1 · p. 2
• DensePed-Lite adapts to observation quality through uncertainty-sensitive scoring, gated complementary scanning paths, and CTDM’s stabilize-verify-refine fusion loop.The mechanisms are designed to reinforce one another rather than act independently.Source: §1 · p. 2
• Experiments on CityPersons and CrowdHuman show a strong accuracy-efficiency trade-off against recent lightweight methods, and comprehensive ablations validate the mechanisms’ coupled effects.Source: §1 · p. 2

2.1 Occlusion Handling

• Adaptive NMS, attention, and multi-modal fusion address occlusion, but reliance on local features or limited semantic context can constrain robustness under heavy occlusion.Source: §2.1 · p. 3

2.2 Multi-scale Modeling

• BiFPN, YOLOv8-CB, and joint modeling advance multi-scale representation, while feature misalignment and boundary ambiguity remain accuracy challenges.Source: §2.2 · p. 3

2.3 Lightweight Detection Methods

• YOLO variants, specialized methods, and DETR-based approaches improve detection modeling, but specialized methods increase cost and DETR-based approaches require more parameters.Source: §2.3 · p. 3

3.1 Overall Framework

• DensePed-Lite targets overconfident localization, fragmented spatial features, and fixed-path information loss through UQE, MPSC, and CTDM.The framework is implemented on YOLOv11n, with MPSC at P4/P5, CTDM integrated with SPPF, and UQE replacing the standard detection head.Source: §3.1 · pp. 3–4

3.2 UQE: Uncertainty-Aware Quality Estimation

• UQE incorporates localization variance as a learnable penalty in the detection score, addressing confidence that can remain high when a box is uncertain.TopK responses alone capture distribution sharpness but not spread; depthwise separable convolutions keep the head efficient.Source: §3.2 · p. 5

3.3 MPSC: Multi-Path Spatial Completion

• MPSC combines complementary scanning paths with adaptive gating to restore spatial coverage when occlusion blocks a path.It is deployed at P4/P5, where 20×20 feature maps reduce overhead compared with 80×80 maps.Source: §3.3 · p. 6

3.4 CTDM: Multi-Scale Coherence Preservation

• CTDM stabilizes feature distributions, verifies global consistency, and refines local details, with residual connections returning refined features to stabilization.The loop lets clear-scale signals guide refinement of ambiguous features while keeping attention lightweight for real-time use.Source: §3.4 · pp. 6–7

3.5 Model Complexity Analysis

• Depthwise separable convolutions reduce UQE complexity, while MPSC and CTDM add controlled overhead and support competitive accuracy with lower computational requirements.MPSC and CTDM are deployed at lower-resolution feature maps.Source: §3.5 · p. 7

4.1 Experimental Setting

• The evaluation uses CityPersons and CrowdHuman, reporting AP50, AP, Recall, Precision, GFLOPs, and parameter count.Training results are averaged over three runs.Source: Datasets · p. 8, Metrics · p. 8, Implementation Details · p. 8

4.2 Experimental Results

• 80.9% AP50 and 50.3% AP on CrowdHuman exceed DEIM-D-FINE-N by 0.8% and 0.5%, respectively, using 5.5 GFLOPs and 2.53M parameters.CityPersons results are 62.1% AP50 and 38.6% AP; Recall rises from 51.3% to 52.9% and Precision from 76.0% to 79.7%.Source: Comparison with Lightweight Detectors · p. 9
• Among ultra-lightweight detectors, DensePed-Lite sits on the Pareto-optimal frontier, with the highest AP50 (80.9%) and second-lowest GFLOPs (5.5).Figure 4 depicts the efficiency-accuracy trade-off on CrowdHuman; bubble size represents parameter count.Source: Efficiency-Accuracy Trade-off · p. 9, Figure 4 · p. 10

4.3 Ablation Studies

• The complete MPSC–UQE–CTDM configuration reaches 62.1% AP50 and 38.6% AP, improving over baseline by 2.5% while reducing GFLOPs by 12.7%.Two-module variants cluster at 61.2–61.5% AP50, whereas the full model reaches 62.1%.Source: Module Combination Analysis · p. 10, Table 2 · p. 11
• The CTDM stabilize-verify-refine loop reaches 60.5% AP50 and 79.9% Precision, while its individual components provide only modest gains in isolation.The reported precision gain is 3.9 points; the loop uses stabilization, verification, and refinement as coupled stages.Source: CTDM Module Analysis · pp. 11–12, Table 3 · p. 11
• Full UQE with learnable λ reaches 61.3% AP50, and Precision improves monotonically from 76.0% to 77.6% across the reported variants.TopK pooling improves baseline AP50 by 0.8%, while adding variance without learnable weighting adds another 0.4%.Source: UQE Uncertainty Modeling Analysis · p. 12, Table 4 · p. 12
• Adaptive gating brings MPSC to 60.2% AP50 and recovers Precision after equal-weight dual paths slightly reduce it.Equal-weight dual paths add only 0.3% AP50 and shift Precision from 76.0% to 75.8%.Source: MPSC Path Redundancy Analysis · pp. 12–13, Table 5 · p. 12

4.4 Visualization Analysis

• Figure 5 shows more concentrated responses for DensePed-Lite than the baseline and DEIM-D-FINE-N on heavily occluded pedestrians and partially visible targets.The reported examples include overlapping pedestrians on CrowdHuman and targets at image boundaries or behind obstacles on CityPersons.Source: §4.4 · p. 13, Figure 5 · p. 13

• On CrowdHuman and CityPersons, DensePed-Lite achieves the best accuracy among ultra-lightweight detectors using 5.5 GFLOPs and 2.53M parameters.Its gains are concentrated in the high-Recall, occlusion-heavy regime the method targets.Source: §5 · p. 14
• The three mechanisms are most effective together, and removing any one leaves a residual failure mode that limits the others.Source: §5 · p. 14
• DensePed-Lite applies quality-aware scoring, multi-path spatial completion, and a stabilize–verify–refine fusion loop to unreliable confidence, fragmented coverage, and incoherent multi-scale fusion.Source: §5 · pp. 13–14
• Evaluation covers two pedestrian benchmarks; broader validation on other crowded categories and embedded hardware is still needed to establish practicality.The quality-aware principle is not specific to pedestrians, and extending it to general dense-object detection is a proposed next step.Source: §5 · p. 14