自動駕駛科技隨近年來人工智慧之快速發展，成為備受矚目的話題，在應用技術上，自駕技術之安全性與即時性成為自駕車系統創新研究的重點之一，而即時感知系統用於確認車輛自身及周邊環境狀態，關係車輛行駛之安全性，故精準且有效地感測周遭環境是實現全自動無人駕駛的核心關鍵技術。
根據文獻得知，多目標追蹤於自駕車感知系統之安全性有重大影響，昔知基於物理模型之感知法則無法處理非線性及非高斯雜訊之感知器數據，物理模型法在執行速度上雖佔有優勢，但無法針對多變場景實際應用；相對地，較複雜之深度學習模型則因計算量龐大，而受限於無法進行即時計算及執行條件。
本論文針對上述問題提出變分自編碼器追蹤方法，在融合物理模型與
深度學習模型上，提出解決上述方法之模型自身限制而產生的固有問題，採用設計合理之權重融合法則，取得處理速度及準確度之間適當平衡，故經驗證後，確認可應用於自駕車多目標追蹤問題，並提供有效且具執行力之方法。

With the rapid advancement of artificial intelligence in recent years, autonomous vehicles have become as a focal point of research and development. Safety and real-time performance
are critical considerations in the design of autonomous driving systems, with perception systems playing a pivotal role in ensuring driving safety by accurately detecting the vehicle's state and its surrounding environment. Precise and efficient perception is fundamental to achieving fully autonomous driving.
Multi-target tracking significantly impacts the safety of autonomous vehicle perception systems. Traditional physics-based methods struggle to handle nonlinear dynamics and non
Gaussian noise, and while they offer advantages in processing speed, they often fail to adapt to complex and dynamic scenarios. Conversely, more sophisticated deep learning models, despite their enhanced accuracy, are computationally intensive and often unable to meet real-time requirements.
This study aims to propose a variational autoencoder-based tracking approach and a welldesigned fusion strategy. By integrating physics-based models with deep learning techniques,
it seeks to address the inherent limitations of both methods. This research strives to achieve an optimal balance between processing speed and accuracy, thereby providing an effective and practical multi-target tracking solution for autonomous driving applications.