Point Cloud Depth Prediction with Adaptive Lists to Reduce Outliers

The patent segments the prediction process by creating multiple candidate lists (first candidate list from same laser, second from different lasers, third from historical points, fourth from priori information) instead of using a single neighboring point prediction. This segmentation allows selective combination of predictions from different sources to handle scenario discontinuity while maintaining manageable complexity through structured organization of candidate predictions.

The patent creates a composite prediction value by combining predictions from multiple candidate lists using weighted summation. The final prediction is composed of elements from different sources (same laser, different lasers, historical data, priori information) with adaptive weights, analogous to composite materials combining different substances to achieve superior properties that single-source prediction cannot provide.

The patent implements dynamic adaptation by updating weights of candidate lists based on prediction residuals. The system dynamically adjusts the importance of each candidate list (same laser, different lasers, historical, priori) according to actual prediction performance, making the prediction structure adaptive rather than static. This dynamic weighting mechanism improves accuracy while managing complexity through performance-driven adaptation.

The patent incorporates feedback loops where prediction residuals are calculated and used to update the weighting of candidate lists. The system continuously monitors prediction accuracy and adjusts the contribution of each candidate list based on this feedback, creating a self-improving prediction mechanism that balances accuracy improvement with complexity management through learned adaptation.