Skip to main content

Wristband Detection

The wristband detection system will provide automated identification of athlete divisions through computer vision analysis of color-coded wristbands worn by all HYROX participants. This component will classify athletes into one of five distinct competition categories (Open Men, Pro Men, Open Women, Pro Women, and Adaptive) by extracting wrist regions using pose estimation keypoints and applying CNN-based color classification, enabling appropriate weight and repetition standards to be applied automatically.

Division Classification Requirements​

HYROX Competition Divisions​

The system must accurately identify and classify athletes into five distinct divisions that determine competition standards:

Open Women - Standard competition category for female athletes using 6kg (14lbs) wall balls with 75 repetitions required. This division represents the majority of female participants and uses standard competition movement patterns without professional-level performance requirements.

Pro Women - Elite competition category for female athletes using 6kg (14lbs) wall balls with 100 repetitions required. Professional athletes in this division compete for prize money and championship titles with increased repetition requirements.

Open Men - Standard competition category for male athletes using 9kg (20lbs) wall balls with 100 repetitions required. This division encompasses the largest participant group with standard competition requirements accessible to recreational athletes.

Pro Men - Elite competition category for male athletes using 9kg (20lbs) wall balls with 100 repetitions required. Professional male athletes compete at the highest level with identical weight requirements but elevated performance standards.

Adaptive Division - Modified competition category for athletes with physical adaptations or movement limitations. This division maintains competitive integrity while allowing customized movement standards and equipment modifications based on individual athlete needs.

Wristband Visual Design​

HYROX Wristband Characteristics​

HYROX wristbands are event-style Tyvek or fabric bands featuring distinctive visual elements that enable reliable computer vision detection. The wristbands display prominent "HYROX" text repeated along the band length and utilize division-specific color schemes.

Color Differentiation serves as the primary detection mechanism, utilizing distinct base colors for each division category. Based on observed samples, divisions may use white/light gray for standard divisions, yellow for specific categories, and potentially other high-visibility colors. Each color maintains sufficient contrast to ensure reliable detection under varied lighting conditions.

Text Elements include "HYROX" branding repeated along the band, though the text size is typically too small for reliable OCR detection from camera distances of 2.1m. The text primarily serves as visual confirmation for human judges rather than automated detection.

Additional Markers may include barcodes or QR codes, but these are not viable for automated reading due to camera distance, oblique viewing angles, and motion during exercises.

Material and Visibility Properties​

Competition wristbands utilize Tyvek or similar synthetic materials that maintain structural integrity and visual clarity throughout multi-day events. The material's matte finish reduces specular reflections that could interfere with computer vision detection while maintaining sufficient brightness for color classification.

Text Contrast Optimization ensures the black "HYROX" branding maintains maximum contrast against all background colors. The bold sans-serif typography is specifically chosen for OCR reliability, with consistent character spacing that aids in text segmentation and recognition.

Computer Vision Detection Pipeline​

Wrist Region Localization​

The detection pipeline leverages the existing pose estimation output to identify wristband locations efficiently. The system extracts wrist joint coordinates directly from the 2D pose estimation pipeline (RTMPose-m with 17 COCO keypoints) and generates a region of interest (ROI) around each wrist.

Pose Keypoint Integration utilizes the wrist keypoints (COCO points 9 and 10) already computed at 90 FPS per camera. This approach eliminates the need for additional object detection models, maintaining the system's sub-200ms latency requirement.

Dynamic ROI Extraction creates a bounding box centered on each wrist keypoint, sized proportionally to the detected arm length (calculated from shoulder-to-wrist distance). The algorithm expands the region by 20-30% to accommodate wristband positioning variability while minimizing background pixels.

CNN-Based Color Classification​

The primary detection mechanism employs a lightweight convolutional neural network (CNN) optimized for color classification of the extracted wrist ROI. This approach provides robust division identification despite varying lighting conditions and wristband orientations.

Neural Network Architecture uses a compact CNN model (similar to MobileNet or ResNet18) fine-tuned specifically for five-way classification of HYROX wristband colors. The model processes 64x64 pixel ROIs extracted from the wrist regions, achieving less than 25ms inference time per frame.

Color Feature Extraction combines learned CNN features with traditional color space analysis in HSV (Hue, Saturation, Value) for improved robustness. The dual approach ensures accurate classification even under challenging venue lighting with shadows and spotlights.

Division Mapping​

The CNN classifier outputs probabilities for each of the five division categories, which are then validated and refined through temporal consistency checks.

Five-Way Classification directly maps wristband colors to competition divisions:

  • Open Women (specific color pattern)
  • Pro Women (specific color pattern)
  • Open Men (specific color pattern)
  • Pro Men (specific color pattern)
  • Adaptive Division (specific color pattern)

Confidence Scoring applies softmax activation to generate classification probabilities. The system requires >80% confidence for immediate classification, with lower confidence scores triggering temporal aggregation over multiple frames.

Temporal Consistency Validation​

Classification decisions incorporate temporal smoothing to prevent sporadic misclassifications during rapid movement or temporary occlusions. The system maintains a rolling buffer of recent classifications to ensure stable division assignment.

Majority Voting aggregates classifications across 15-20 consecutive frames (approximately 250-330ms at 60fps) before confirming division assignment. This approach filters out transient errors from motion blur or partial occlusions.

Confidence Thresholding prevents low-confidence classifications from affecting athlete division assignment. Classifications below 80% confidence trigger extended temporal aggregation or prompt manual verification by competition judges.

Classification Architecture​

Lightweight CNN Model​

The classification system employs a purpose-built CNN optimized for real-time inference on edge devices. The model architecture balances accuracy with latency constraints to maintain the system's sub-200ms end-to-end processing requirement.

Model Architecture uses a compact design with:

  • Input layer: 64x64x3 RGB images from wrist ROIs
  • Convolutional layers: 3-4 blocks with batch normalization and ReLU activation
  • Global average pooling to reduce parameters
  • Output layer: 5-way softmax classification
  • Total parameters: Less than 1M for edge deployment efficiency

Training Strategy employs:

  • Transfer learning from ImageNet-pretrained backbones
  • Data augmentation for lighting and color variations
  • Hard negative mining for challenging lighting conditions
  • Validation on competition venue footage

Fallback Mechanisms​

Secondary Classification Methods​

When wristband detection fails due to occlusion, absence, or damage, the system employs alternative classification strategies to maintain operational continuity. These fallback mechanisms prioritize competition flow while ensuring accurate division assignment.

Manual Override Interface enables competition judges to input division information directly when automated detection is unavailable. The interface provides rapid division selection with visual confirmation to prevent data entry errors.

Quality Assurance Protocols​

Comprehensive validation ensures classification accuracy throughout competition events. The system continuously monitors detection performance and alerts operators when intervention may be required.

Detection Quality Metrics track classification confidence and temporal consistency across frames. Performance degradation (confidence below 70% sustained) triggers proactive alerts to technical staff for system adjustment or manual intervention.

Audit Logging maintains detailed records of all classification decisions, including confidence scores, detection modalities used, and any manual overrides applied. These logs support post-event analysis and dispute resolution procedures.

Integration with Competition Systems​

Real-Time Division Application​

Detected division classifications immediately influence competition parameters including weight selection, repetition counting, and performance standards. The system seamlessly communicates division information to all connected judging components.

Weight Standard Selection automatically configures the system to expect appropriate wall ball weights based on detected division. Male divisions (Open Men, Pro Men) trigger 9kg weight validation while female divisions (Open Women, Pro Women) expect 6kg implements.

Repetition Target Adjustment modifies rep counting thresholds based on division requirements. Pro Women athletes must complete 100 repetitions compared to 75 for Open Women, while both male divisions require 100 repetitions.

Judge Interface Integration​

The wristband recognition system provides clear division status information to competition judges through intuitive visual interfaces. Real-time displays show current athlete division, detection confidence, and any system alerts requiring attention.

Visual Status Indicators use color-coding and iconography consistent with wristband designs to minimize cognitive load on judges. Division status appears prominently on judge displays with confidence metrics and timestamp information.

Override Capabilities allow judges to correct or confirm division assignments when necessary. The interface requires deliberate confirmation actions to prevent accidental changes while maintaining the flexibility needed for adaptive athlete accommodations.

Privacy and Compliance​

Data Minimization​

The wristband recognition system processes only the minimal visual information necessary for division classification. The system explicitly avoids capturing or storing biometric data, facial features, or other personally identifiable information.

Focused Processing restricts analysis to wristband regions without processing or storing broader body imagery. Temporary image buffers are immediately purged after classification, preventing accumulation of visual data that could raise privacy concerns.

Anonymous Classification assigns divisions without linking to specific athlete identities. The system maintains functional separation between division classification and athlete identification systems to ensure privacy protection.

Regulatory Compliance​

System design ensures compliance with international privacy regulations including GDPR, CCPA, and other regional frameworks. Privacy-by-design principles guide all architectural decisions to prevent regulatory conflicts.

Transparent Operation provides clear information to athletes about wristband detection purposes and methods. Competition registration materials explain the automated division classification system without requiring explicit consent for non-biometric visual processing.

Data Protection Measures implement comprehensive security controls to prevent unauthorized access to classification systems. All processing occurs locally without cloud transmission, reducing privacy risks associated with data transfer and remote storage.

Real-World Detection Challenges​

Wristband Positioning Variability​

Athletes wear wristbands at different positions and orientations, creating detection challenges that require robust algorithmic solutions. Wristbands may be rotated around the wrist, partially covered by clothing, or worn at varying distances from the wrist joint.

Rotation Invariance is achieved through CNN training with extensive data augmentation including rotations, flips, and perspective transforms. The model learns color features that are robust to wristband orientation.

Partial Occlusion Handling leverages temporal consistency across frames. When the wristband is temporarily obscured, the system maintains the last high-confidence classification for up to 2 seconds before requesting manual verification.

Lighting and Motion Artifacts​

Competition venues present challenging lighting conditions with dramatic shadows, bright spotlights, and rapid lighting changes. Additionally, the fast-paced nature of wall ball exercises creates motion blur that can compromise detection quality.

Motion Blur Compensation is handled through the global shutter cameras (IMX273 sensors) already deployed for pose estimation. The 60-120 fps capture rate with short exposure times minimizes motion artifacts during rapid arm movements.

Dynamic Range Optimization leverages the CNN's training on diverse lighting conditions. The model has been trained on synthetic and real data covering various lighting scenarios, from dark warehouse venues to bright television lighting setups.

Performance Optimization​

Real-Time Processing Requirements​

The classification pipeline achieves less than 25ms latency from ROI extraction to division determination through careful optimization. The lightweight CNN model and efficient preprocessing ensure consistent performance across multiple simultaneous athletes.

GPU Acceleration utilizes TensorRT optimization for the CNN inference, achieving FP16 precision with minimal accuracy loss. The same GPU infrastructure used for pose estimation handles division detection without additional hardware requirements.

Pipeline Integration processes division detection in parallel with other inference tasks. While pose estimation runs on the main frame, division detection operates on the extracted ROIs, maximizing GPU utilization without blocking the critical path.

Scalability Considerations​

System architecture supports simultaneous classification across 40-80 wall ball stations typical of major HYROX events. Distributed processing capabilities ensure consistent performance regardless of venue size or athlete density.

Load Balancing distributes classification workload across available computing resources based on real-time demand. Dynamic resource allocation prevents performance degradation during peak competition periods with maximum athlete density.

Caching Strategies store recent classification results to reduce redundant processing for stationary athletes. Intelligent cache invalidation ensures updates when athletes change positions while minimizing unnecessary recomputation.