摘要 準確預測行人軌跡,是自主系統在與人類共享的環境中運作時的一項基本需求。本論文提出 人類軌跡預測模型(Human Trajectory Prediction Model, HTPM),這是一種用於多智能體行人 軌跡預測的機率式潛在變數架構。HTPM 使用具有時間注意力池化的長短期記憶網路(Long Short-Term Memory, LSTM)局部時間編碼器,對每一位智能體的觀測運動歷史進行編碼。情境 投影層將所得的智能體嵌入映射為精簡表示,而輕量化全域聚合器則透過跨智能體的注意力池 化形成場景層級描述子。條件式潛在模組參數化一個對角高斯先驗,並在訓練期間參數化一個 變分對角高斯後驗。取樣得到的潛在碼由自迴歸門控循環單元(Gated Recurrent Unit, GRU)解 碼器進行解碼;該解碼器以涵蓋完整預測時間範圍的多步負對數似然目標進行訓練。 本模型在 ETH-UCY 基準上進行評估,使用五個標準行人場景:ETH、Hotel、Univ、Zara-1 與 Zara-2。在留一法(leave-one-out)best-of-20 協定下,HTPM 在五個資料分割上的平均 ADE 為 0.1174 m,平均 FDE 為 0.1969 m;場景層級 ADE 的範圍從 Zara-2 的 0.099 m 到 ETH 的 0.145 m。另於 Stanford Drone Dataset(SDD)上進行額外評估,並採用 PECNet/Y-Net 的像素 空間前處理慣例(÷20),得到 ADE = 5.97 ± 0.03 pixels 與 FDE = 9.12 ± 0.08 pixels,結果來自 100 次獨立測試執行。這些結果顯示,在所述前處理設定下,隨機取樣具有穩定性;然而,更 廣泛的跨資料集泛化仍取決於資料集尺度與攝影機幾何條件。 本論文亦提出一套運動感知視覺化流程,結合 YOLOv12x 行人偵測、BYTETracker 身分追 蹤與 HTPM 軌跡預測。三層級運動分類器會對靜止智能體抑制預測扇形,對慢速移動智能體繪 製精簡扇形,並對步行智能體顯示五假設扇形。ETH Hotel 與 Zara-1 上的定性範例顯示,在所 選影格中,該方法能減少視覺雜訊,並改善預測扇形與觀測運動方向之間的對齊。HTPM 含有 516 K 個參數,需要 28 M 次乘加運算(multiply-accumulate operations, MACs),並在 NVIDIA RTX 3090 GPU 上達到每位智能體平均 23 µs 的推論延遲,支持即時桌上型 GPU 推論,並為未 來與機器人導航系統整合提供動機。 關鍵詞:軌跡預測、行人運動、潛在變數模型、變分推論、LSTM、GRU、ETH-UCY、Stanford Drone Dataset、BYTETracker、運動分類、自主系統。;Abstract Accurate prediction of pedestrian trajectories is a fundamental requirement for autonomous systems operating in shared human environments. This thesis presents the Human Trajectory Prediction Model (HTPM), a probabilistic latent-variable architecture for multi-agent pedestrian trajectory prediction. HTPM encodes each agent’s observed motion history using a Long Short- Term Memory (LSTM) local temporal encoder with temporal attention pooling. A context projection layer maps the resulting agent embedding into a compact representation, and a lightweight global aggregator forms a scene-level descriptor through attentive pooling across agents. A conditional latent module parameterises a diagonal Gaussian prior and, during training, a variational diagonal Gaussian posterior. Sampled latent codes are decoded by an autoregressive Gated Recurrent Unit (GRU) decoder trained with a multi-step negative log-likelihood objective over the full prediction horizon. The model is evaluated on the ETH-UCY benchmark using the five standard pedestrian scenes: ETH, Hotel, Univ, Zara-1, and Zara-2. Under the leave-one-out best-of-20 protocol, HTPM obtains a mean ADE of 0.1174 m and a mean FDE of 0.1969 m across the five splits, with scene- level ADE ranging from 0.099 m on Zara-2 to 0.145 m on ETH. Additional evaluation on the Stanford Drone Dataset (SDD), using the PECNet/Y-Net pixel-space preprocessing convention (÷20), gives ADE = 5.97 ± 0.03 pixels and FDE = 9.12 ± 0.08 pixels over 100 independent test runs. These results indicate stable stochastic sampling under the stated preprocessing settings, while broader cross-dataset generalisation remains dependent on dataset scale and camera geometry. The thesis also introduces a motion-aware visualisation pipeline that combines YOLOv12x pedestrian detection, BYTETracker identity tracking, and HTPM trajectory prediction. A three- tier motion classifier suppresses prediction fans for stationary agents, renders compact fans for slow-moving agents, and displays five-hypothesis fans for walking agents. Qualitative examples on ETH Hotel and Zara-1 illustrate reduced visual clutter and improved alignment between predicted fans and observed motion direction in selected frames. HTPM contains 516 K parameters, requires 28 M multiply-accumulate operations (MACs), and achieves a mean inference latency of 23 µs per agent on an NVIDIA RTX 3090 GPU, supporting real-time desktop-GPU inference and motivating future integration with robotic navigation systems. Keywords: Trajectory Prediction, Pedestrian Motion, Latent Variable Model, Variational Infer- ence, LSTM, GRU, ETH-UCY, Stanford Drone Dataset, BYTETracker, Motion Classification, Autonomous Systems.