Probability-Based Map Fusion — Research Paper

SLAM-based robotic mapping systems produce geometric representations of the environment but lack semantic understanding of what objects are present and where. Bridging this gap by incorporating neural network object labels into the map calls for a lightweight probabilistic approach that remains tractable on resource-constrained robotic hardware.

Developed a probability-based fusion algorithm that aggregates neural network object detections over time, weighting observations by distance from stored map points using radial basis functions and the midpoint circle algorithm for efficient spatial cell filtering. The method produces a probabilistic object map without requiring expensive re-computation.

Demonstrated that lightweight, probability-based methods can produce richer, context-aware maps that support object search, counting, and location estimation — improving robot decision-making without sacrificing performance.

• •Probability-based aggregation of object detections over time
• •Weighted distance using radial basis functions for spatial inference
• •Midpoint circle algorithm for efficient occupancy cell traversal
• •Fuses SLAM occupancy grid data with CNN object detection labels (YOLO, EfficientNet, MobileNet)
• •Supports searchable object databases and occurrence estimation in mapped regions
• •Designed for onboard efficiency on resource-constrained robotic hardware