How We Enabled Autonomous Navigation For Golf Carts In Unstructured Campus Environments

The client wanted to develop an ADAS (Advanced Driver Assistance System) for a golf cart operating autonomously on a campus with limited infrastructure. Unlike conventional roadways, the terrain featured unmarked paths, irregular surfaces, and lacked road signs or lane markers, posing significant challenges for conventional perception and planning models.

As the project transitioned from initial prototypes toward real-world testing, the team encountered critical roadblocks in achieving consistent performance, safety, and navigational intelligence.

So, the main objectives were to:

• Develop a perception system capable of handling unstructured environments with limited or no lane markings, ensuring robust road understanding without relying on traditional lane detection models.
• Enhance path planning accuracy and smoothness to maintain passenger comfort and ensure safe turns in environments with sharp curves and narrow lanes.
• Achieve consistent and precise steering control by reducing reliance on noisy or inconsistent sensor readings and introducing more intelligent steering logic.
• Improve safety with dynamic obstacle awareness, allowing the system to detect and react to unexpected pedestrians, vehicles, or objects on the drivable path in real time.
• Create a scalable and modular ADAS architecture that can be generalized to other low-infrastructure or semi-structured environments such as campuses, resorts, and industrial zones.

To meet these objectives, the team needed a tailored perception and controlled pipeline capable of intelligent environment interpretation, smooth path generation, and real-time decision-making, independent of well-structured road inputs.

Our engineering team conducted a thorough assessment of the existing perception and control pipeline and identified multiple system limitations. The ADAS system was initially reliant on conventional lane detection models, which underperformed in unmarked terrains, resulting in erratic path estimation and unreliable steering.

Before Implementation

• Lane detection models underperformed due to a lack of clear markings, affecting path generation.
• Vision segmentation algorithms failed to define road boundaries, leading to inconsistent trajectory outputs.
• Steering logic depended heavily on noisy sensor feedback, causing jerky and unstable movements.
• Static obstacle avoidance failed to react to dynamic environmental changes, creating potential safety concerns.

After Implementation

• Introduced a custom perception model using pole and boundary detection for virtual lane estimation, enabling accurate centreline estimation even in unstructured settings.
• Applied path smoothing algorithms like Bezier curves and polynomial fitting with frame averaging to create smooth, continuous driving paths.
• Applied path smoothing algorithms using polynomial fitting and Bezier curves, supported by frame-level averaging, to create continuous, jitter-free trajectories.
• Replaced raw steering angle input with a radius-of-curvature-based control logic significantly improving accuracy and driving smoothness across varied road structures.
• Integrated multi-sensor fusion—GPS, IMU, proximity sensors, and cameras—to allow dynamic obstacle detection and autonomous rerouting, increasing safety and reliability.

These enhancements transitioned the system from a sensor-limited, reactive platform to a context-aware, intelligent ADAS capable of safely navigating real-world campus roads.

Key Highlights

• Custom Lane Detection via Environmental Feature Mapping: Developed a vision-based segmentation model to detect poles and curbs, enabling virtual lane estimation in the absence of standard road markings.
• Smoothed Trajectory Generation: Implemented Bezier curve and polynomial path smoothing with frame-level averaging for accurate, jitter-free path planning.
• Curvature-Based Steering Control: Replaced traditional steering logic with a curvature-driven model to achieve smoother, more stable path following.
• Dynamic Obstacle-Aware Path Rerouting: Integrated real-time obstacle detection and rerouting to ensure safety and uninterrupted autonomous navigation.
• Sensor Fusion for Precision Localization: Combined inputs from GPS, IMU, LiDAR, and cameras to enhance real-time localization, path correction, and decision-making accuracy.

• 65% Improvement in Path Prediction Accuracy
• 50% Improvement in Steering Smoothness and Control
• Real-Time Obstacle Response Achieved
• Improved Navigation Reliability in Low-Structure Environments