Lightweight YOLOv8 with Multi-Module Optimization for Electric Bicycle Rider Helmet Detection

Technical Summary: Lightweight YOLOv8 with Multi-Module Optimization for Electric Bicycle Rider Helmet Detection

Problem Statement
The rapid proliferation of electric bicycles has introduced significant traffic safety challenges, primarily due to non-standard helmet-wearing behaviors. While mandatory helmet regulations exist, traditional manual supervision is inefficient and lacks full coverage. Existing deep learning-based detection algorithms face critical bottlenecks in real-world deployment:

1. Small Object Detection: Difficulty in accurately detecting distant or small-scale helmet targets.
2. Environmental Robustness: Degraded performance under complex backgrounds, occlusion in dense traffic, and varying lighting conditions.
3. Computational Efficiency: Many models utilize bloated network structures that are computationally expensive, hindering deployment on mobile and edge devices required for real-time traffic management.
4. Balancing Act: A lack of effective trade-offs between detection accuracy, real-time inference speed, and model lightweightness.

Methodology
The authors propose an improved YOLOv8-based detection system designed to enhance feature extraction for small targets and improve adaptability to complex traffic scenarios. The baseline model selected is YOLOv8s, chosen for its balance between accuracy and speed. The proposed method integrates four specific multi-module optimizations:

1. Loss Function Optimization (Inner-MPDIoU):

   • The original CIoU loss is replaced with Inner-MPDIoU (Inner Multi-Point Distance Intersection over Union).
   • Mechanism: Inner-MPDIoU minimizes the distance between the predicted and ground truth box corner vertices ($d_1, d_2$) while introducing a ratio factor to control an auxiliary bounding box.
   • Benefit: This approach reduces computational complexity by eliminating inverse trigonometric functions, improves convergence stability, and specifically enhances localization accuracy for small targets and targets with unbalanced aspect ratios.

2. Backbone Enhancement (C2f_CA Module):

   • The standard C2f module in the YOLOv8 Backbone is replaced with a C2f_CA module, which integrates the Coordinate Attention (CA) mechanism.
   • Mechanism: CA performs one-dimensional global pooling along horizontal and vertical directions to encode spatial position information.
   • Benefit: This allows the model to adaptively filter irrelevant background noise and highlight key target features (helmets/electric bikes) without significantly increasing computational overhead, improving focus on dense and small targets.

3. Small Object Detection Layer Addition:

   • A dedicated 160×160 detection layer is added to the Neck structure, alongside the standard 20×20, 40×40, and 80×80 layers.
   • Mechanism: This layer utilizes high-resolution feature maps to capture fine-grained details often lost during downsampling.
   • Benefit: It significantly reduces missed detections and false alarms for ultra-small targets (e.g., distant helmets) by fusing shallow and deep semantic information.

4. Upsampling Optimization (CARAFE):

   • Traditional upsampling methods (nearest-neighbor or bilinear) are replaced with CARAFE (Content-Aware Re-Assembly of Features).
   • Mechanism: CARAFE uses a Kernel Prediction Module (KPM) and Content-Aware Reassembly Module (CARM) to generate dynamic weights based on input feature content.
   • Benefit: This preserves local details and semantic information during feature fusion, enhancing the quality of upsampled feature maps for small object detection.

Key Contributions

• Multi-Module Synergy: The paper presents a holistic optimization strategy combining loss function refinement, attention mechanisms, architectural expansion for small objects, and advanced upsampling, rather than relying on a single modification.
• Lightweight Efficiency: Despite adding modules, the model maintains a parameter count (3.14M) and computational cost (13.6 GFLOPs) comparable to other lightweight variants, ensuring suitability for edge deployment.
• Targeted Small-Object Optimization: The specific addition of the 160×160 layer and the use of Inner-MPDIoU directly address the specific challenge of detecting small, distant helmets in traffic.

Experimental Results
The study was conducted using two datasets (1,500 and 1,200 images) sourced from Roboflow, split 8:2 for training and testing. The environment utilized an NVIDIA RTX 3060 GPU.

• Baseline Comparison: The proposed model outperformed other lightweight YOLO variants (v5, v6, v9, v10, v11) in terms of mAP50 and mAP50-95 while maintaining competitive inference speeds.
• Ablation/Comparative Analysis: The proposed method ("Ours") was compared against three other improved algorithms:
  • YOLOv8-otl (CARAFE + WIoU)
  • YOLOv8-C2fDCN (EMA + Inner-MPDIoU)
  • YOLOv5-CA (CA + CBAM + DIoU)
• Performance Metrics:
  • On Dataset 1, the proposed method achieved the highest mAP50 (0.638) and mAP50-95 (0.234), with a Precision of 0.652 and Recall of 0.654.
  • On Dataset 2, it achieved an mAP50 of 0.879 (tying the best) and the highest mAP50-95 (0.612).
  • The model demonstrated superior classification accuracy in confusion matrix analysis, with higher diagonal values (correct classifications) and lower off-diagonal values (misclassifications) compared to competitors.
  • The Precision-Recall (P-R) curve for the proposed method was consistently higher, indicating a larger Area Under the Curve (AP).

Significance and Claims
The paper claims that the proposed method successfully addresses the trade-off between detection accuracy and real-time performance in electric bicycle helmet detection. By effectively suppressing false positives and missed detections in complex traffic scenarios, the system provides a reliable technical solution for intelligent traffic safety management. The authors assert that their multi-module collaborative optimization strategy offers a more practical and robust approach for edge device deployment compared to existing single-module improvements. The study concludes that this algorithm provides a feasible foundation for automated, large-scale supervision of helmet-wearing behaviors, though future work is suggested to expand dataset diversity and explore further model compression for extreme environments.