Lightweight Edge-Based Telemetry Parsing for Navigation Systems via Deterministic Spatial-Text (LODTE)

Technical Summary: Lightweight Edge-Based Telemetry Parsing for Navigation Systems via Deterministic Spatial-Text (LODTE)

Problem Statement
Traditional methods for parsing real-time navigation telemetry (such as Estimated Time of Arrival and distance metrics) from digital map interfaces rely heavily on cloud-centric vision architectures or resource-intensive Convolutional Neural Networks (CNNs). The authors identify three critical limitations in these existing approaches:

1. Network Latency: Round-trip transmission to cloud servers renders Optical Character Recognition (OCR) impractical for sub-second, safety-critical edge responses.
2. Bandwidth Overhead: Continuous video streaming required for cloud processing demands significant network throughput.
3. Privacy Risks: Transmitting sensitive screen or cabin telemetry off-device introduces data privacy vulnerabilities.
   Furthermore, traditional template-matching approaches often fail under scale or perspective variance, while heavy neural networks strain low-power edge processors.

Methodology: The LODTE Framework
The paper proposes the Lightweight Edge-Based Telemetry Parsing (LODTE) framework, a deterministic, edge-native vision architecture designed to parse navigation UI elements locally without cloud dependency. The system operates via a four-stage pipeline optimized for hardware with limited resources:

1. Edge-Optimized Preprocessing: Input frames captured via high-performance screen buffers (using mss) are converted to grayscale to minimize channel processing overhead. A localized Gaussian blur kernel is applied to suppress high-frequency background noise from map vectors and road networks.
2. Deterministic Spatial-Text Bounding: This stage isolates telemetry cards (e.g., ETA/distance) by applying strict geometric priors rather than learning-based detection. The system filters for bounding boxes with specific dimensions (width: 52–168px; height: 38–54px). To eliminate false positives from map artifacts, a multi-criteria validation filter is applied:
   • Aspect Ratio (AR): Enforces a constraint of $1.5 \le AR \le 3.2$ to distinguish horizontal telemetry banners from square pins or vertical markers.
   • Mean ROI Intensity ($\mu_{ROI}$): Requires a mean pixel intensity $> 210$ to ensure the target is a solid, high-contrast light background typical of active UI cards.
   • Central Score: Penalizes edge anomalies to ensure robust target isolation.
3. Resource-Constrained OCR Tokenization: Isolated regions of interest (ROI) are upsampled using bicubic interpolation (3×) to enhance font edge definition. The data is then passed to an optimized Tesseract OCR engine configured for uniform text blocks (--psm 6), avoiding the overhead of deep-learning inference.
4. Cumulative Latency Profiling: The framework includes a built-in profiler measuring execution time across all stages to ensure real-time performance.

Key Contributions

• Deterministic Spatial-Text Bounding: The introduction of a non-neural, geometry-based approach that enforces strict dimensional and aspect-ratio priors to localize UI elements, removing the need for heavy object detection models.
• Edge-Native Architecture: A complete pipeline that eliminates cloud round-trip overhead, enabling offline operation and preserving data privacy.
• Optimized Tokenization: A lightweight workflow combining 3× bicubic upscaling and adaptive binarization (Otsu's method) to achieve high recognition fidelity with minimal computational power.

Experimental Results
Empirical benchmarking was conducted across various navigation scenarios, including short-range commutes, mid-range regional routes, and long-distance complex routing.

• Latency: The system achieved an average total processing latency of 11.48 ms per frame. In specific test cases, cumulative latency ranged between 10.66 ms and 12.87 ms.
• Stage Breakdown:
  • Preprocessing: ~2.14 ms (6.5% of total time).
  • Spatial Bounding: ~4.85 ms (14.7% of total time).
  • OCR Tokenization: ~25.10 ms (76.2% of total time).
• Accuracy: The framework demonstrated 100% character segmentation accuracy on verified samples with consistent aspect ratios. Extraction success was high for standard UI layouts, though the authors noted that irregularly padded or smaller bounding boxes occasionally introduced OCR noise, leading to partial extraction statuses.
• Robustness: The system maintained performance across varying route complexities and background map densities without requiring cloud infrastructure.

Significance and Claims
The paper claims that LODTE successfully bridges the gap between high-accuracy cloud vision and the latency/privacy constraints of edge computing. By removing cloud dependencies, the framework enables real-time, safety-critical navigation telemetry parsing with microsecond-level efficiency. The authors assert that the system maintains high optical recognition fidelity even when background complexities vary, proving that sub-second latency can be sustained without severely compromising spatial-textual data integrity.

The authors conclude that while the framework is robust for structured UI layouts with consistent aspect ratios, future work should focus on extending UI adaptability for irregularly padded elements and integrating multi-scale feature pyramids to further improve recognition across varied screen resolutions.