Seyond™ utilizes their in-house developed single-laser Forward SightNet and multi-laser fusion perception lightweight model for efficient computation. They utilize time-space fusion 3D object detection and tracking algorithms, and combine traditional methods such as clustering and Kalman filtering to reduce computational load. The algorithm takes into account the balance between real-time and accuracy, and integrates BEV and RV dual-model detectors to provide more precise object boxes at different perception distances.
Seyond has developed an interactive interface for visualizing raw point cloud data and perception results, with three-view windows synchronizing point cloud and image data. It supports real-time execution of point cloud perception algorithms, as well as two offline replay modes. During data collection and storage, road condition information can be tagged, making data management more convenient and friendly to both online testing and offline
Seyond has developed an automatic labeling software tool for laser point cloud data, supporting batch labeling of continuous multiple frames of 3D object boxes and point cloud segmentation. The tool is highly automated, requiring only minimal human verification, and can improve labeling efficiency by dozens of times compared to pure manual labeling. The vast labeling data can be automatically synthesized by algorithms to obtain ground truth and provide convenience for simulation testing.
Seyond has independently developed an evaluation software tool for laser radar perception performance. By setting parameters such as thresholds and confidence levels, the evaluation tool compares the perception output results with ground truth based on single-module, multi-module, and global functions of the perception algorithm. It generates visual data analysis charts and evaluation analysis reports, and provides a cloud service for managing vast laser point cloud data.