Lane Detection Using HCNN and DBSCAN Clustering

The patent segments the lane detection task into multiple processing stages: initial lane line detection, clustering analysis using DBSCAN algorithm, and outlier filtering. This segmentation allows the system to process detection results in discrete steps, improving reliability by systematically handling optical interference at each stage rather than relying on a single vulnerable detection pass.

The patent implements feedback mechanisms where detection results are continuously refined through clustering analysis. The DBSCAN algorithm processes detected lane lines and provides feedback by identifying clustered data points as valid lane lines and outlier data points as noise. This feedback loop continuously improves detection reliability by eliminating false positives caused by optical interference from weather and debris.

The patent segments the neural network architecture into a heterogeneous convolutional neural network with distinct convolution layers, pooling layers, and fully connected layers. Each layer type performs a specific function: convolution layers for feature extraction, pooling layers for dimensionality reduction, and fully connected layers for classification. This segmentation allows the system to achieve high detection accuracy while managing complexity through modular, specialized processing stages.

The patent extracts and applies the DBSCAN clustering algorithm as a separate post-processing step after the neural network generates initial detection results. This extraction of the clustering function from the main neural network pipeline allows the system to achieve high measurement precision by combining neural network feature extraction with specialized clustering analysis, while managing device complexity by separating concerns between the neural network and the clustering algorithm.

The patent extracts and applies the DBSCAN clustering algorithm as a computationally efficient post-processing step that operates on detected lane line data rather than raw images. This extraction allows the system to process data from multiple cameras with wide field of view by working with condensed feature representations rather than full-resolution images, thereby maintaining broad road coverage while improving processing efficiency and reducing computational requirements.

The patent discards redundant and outlier data points identified by the DBSCAN clustering algorithm while preserving and recovering valid lane line information. By filtering out outlier data points that represent noise or false detections, the system reduces the data processing load from multiple cameras while maintaining accurate lane detection across the wide coverage area, thereby improving productivity without sacrificing detection accuracy.