高性價車載影像分析軟體引擎

車電人工智慧化產業技術研發計畫

1. 可偵測/分割物件包含行人、騎士、二輪車、四輪車、車道線、可行駛區域
2. 準確率 ≧90%

雛型

本技術利用單一架構來同時具備物件偵測與語意分割功能，並可以依客戶產品所使用的硬體平台及算力來選擇適用的AI模型，將可協助國內廠商發展具備性價比最佳化的產品，爭取國際市場，潛力可期。

自動駕駛系統、行車環境感知系統

無

深度學習演算法

本技術應用於需要同時具備物件偵測與語意分割雙重任務的AI應用，可偵測行人,騎士, 2輪車, 4輪車, 車道線與可行駛區域. 主要係利用單一(Backbone, 多重任務分支)架構來同時具備物件偵測與語意分割功能。其backbone架構可依平台算力而有不同選擇，例如Yolov8、Yolov9、Yolov10, Yolov11中不同複雜度的n, s, m, l, x的模型架構。

This technology is applied in AI applications that require both object detection and semantic segmentation tasks simultaneously. It is capable of detecting pedestrians, cyclists, two-wheeled vehicles, four-wheeled vehicles, lane markings, and drivable areas. The core of the solution utilizes a unified architecture (Backbone with multi-task branches) that enables both object detection and semantic segmentation functionalities. The backbone architecture can be adapted to various platform computational capabilities, offering different model complexities such as Yolov8, Yolov9, Yolov10, and Yolov11, with corresponding variants including n, s, m, l, and x models based on the desired complexity.