Context: Research Motivation & Environmental Challenges
As part of a 28-page research study for technology initiatives aligned with Qatar National Vision 2030, I built a 1:10 scale autonomous vehicle testbed powered by an NVIDIA Jetson Nano. The project was motivated by environmental challenges relevant to Qatar—such as high ambient light, thermal stress, and dust—while prototype testing was conducted on an indoor obstacle course with simulated high-glare lighting.
The Sensor Dilemma: Budget LiDAR vs. Edge Vision
In hobbyist and academic robotics, 2D spinning LiDARs (such as the RP-LiDAR A1) are frequently recommended as the default obstacle detection sensor. However, lab testing revealed several trade-offs:
1. Surface Absorption: Low-cost infrared LiDAR beams suffered attenuation when encountering matte black obstacles and reflective glossy surfaces common in urban road layouts.
2. Elevation Blindness: 2D planar LiDAR scans only a single horizontal cross-section, missing low-profile road debris or hanging obstacles.
3. Thermal Vulnerability: Continuous motor rotation under warm ambient temperatures requires active thermal monitoring.
The Computer Vision Alternative
To address these challenges, I implemented a wide-angle CSI camera pipeline running TensorRT-quantized YOLOv8 Nano directly on the Jetson Nano GPU (128-core Maxwell architecture).
[CSI Camera Stream] → [V4L2 Frame Capture] → [CUDA BGR to RGB Conversion]
↓
[TensorRT INT8 Engine: YOLOv8n] → [Non-Maximum Suppression (NMS)]
↓
[Spatial Bounding Boxes + Distance Estimator] → [C++ PID Steering Controller]Comparative Observations
| Feature | RP-LiDAR A1 (2D) | YOLOv8n (TensorRT FP16) |
|---|---|---|
| Field of View | 360° Planar (Horizontal only) | Wide-angle (Horizontal + Vertical) |
| Scan / Frame Rate | Fixed mechanical scan rate | Interactive inference frame rate |
| Object Classification | Range-only (No classification) | Semantic classes (Vehicles, Cones, Obstacles) |
| Surface Sensitivity | Dependent on material reflectivity | Robust optical detection across textures |
| Compute Overhead | Low CPU host overhead | GPU-bound inference with active thermal dissipation |
Conclusion
For compact autonomous platforms designed with warm, bright environments in mind, an edge-computed deep neural vision pipeline provides rich spatial intelligence alongside distance sensors, provided proper quantization and thermal dissipation are applied. Prototype testing was performed in controlled indoor lab conditions.