隨著現代科技發展以及人工智慧的崛起，越來越多人關注自動駕駛的相關研究。自駕車技術涵蓋多面向層次，凡如感測器融合、機器學習、路徑規劃及影像辨識皆屬關聯項目，基於自駕車技術能夠改善人類行動安全及輔助車輛運作，更可延伸到其它應用領域之自走機器人產業。
　　本論文研究之目的係針對自駕車模型之代表法則，建構基於中心點架構之 3D 物體檢測及追蹤模型，因應處理大量光達點雲資料運算。主要概念是將 3D 物體中心視為關鍵點，不採用3D邊界框進行計算（Bounding Box），用以更簡化追蹤計算流程，並更容易進行實時（Real-Time）執行。在檢測中心點數據，用來推斷物體其它屬性，例如大小、方向及速度，保留關鍵資訊用於後續追蹤。
　　實驗驗證部分，本論文以nuScenes開放資料集進行評估，分析包含 MOTA、AMOTA、Recall及ID Switch 數量指標，並比較行人、車輛、巴士及腳踏車在不同策略下表現之優劣特性。在不顯著增加運算成本的條件下，本改良方法確可有效降低 ID 切換及誤檢，並提升在高recall條件下的MOTA 成效。故本論文方法提供一套輕量的追蹤系統優化策略，對於自駕系統之三維感知模組具有參考價值。

With the advancement of modern technology and the rise of artificial intelligence, autonomous driving has attracted growing interest. The development of self-driving technology encompasses multiple domains, including sensor fusion, machine learning, path planning, and image recognition. As these technologies can enhance traffic safety and assist vehicle operation, they are also being extended to other fields such as autonomous robotics.
This thesis aims to develop a 3D object detection and tracking system based on a center-based architecture, specifically designed to handle large-scale LiDAR point cloud data. The key concept is to treat the center of a 3D object as a critical point instead of using 3D bounding boxes; thereby, simplifying the tracking computation and enabling real-time performance. Center point predictions are used to infer object attributes such as size, orientation, and velocity, preserving essential information for subsequent tracking.
The proposed method is evaluated on the nuScenes open dataset, using metrics including MOTA, AMOTA, recall, and the number of ID switches. The study further analyzes the performances across various object types- pedestrians, vehicles, buses, and bicycles under different strategies. Without significantly increasing computational cost, the proposed approach effectively reduces ID switches and false detections, and improves MOTA under high recall conditions. Hence, this study provides a lightweight optimization strategy for tracking systems, serving as a reference for 3D perception modules in autonomous driving.