本研究旨在探索適合應用於嵌入式微型無人機之單目視覺避障技術，解決現
有光流演算法在低功耗平台上運算效率不足的問題。
探討三種單目視覺避障技術。第一種為 Simple Lucas-Kanade 光流演算法，
計算量小，可在低效能晶片上運行，但精確度較低。 由於後兩種演算法計算量
較大，先於模擬平台測試，再評估移植至嵌入式平台的可行性。第二種為 DIS
(Dense Inverse Search) 光流演算法，精準度最高但計算量大。第三種為 SIF 演
算法，準確度優於 Simple Lucas-Kanade，計算量小於 DIS，更適合微型無人機。
此外，本研究於模擬環境結合 Flowdep 演算法，將光流轉換為影像深度資訊，
以提升避障效果。
針對無人機避障策略，本研究提出兩種方法。 嵌入式平台上，畫面分為左
右兩半，根據光流變化決定避障方向；模擬環境則採用區域劃分策略，去除不必
要的資訊，以降低計算負擔並確保系統的準確性與穩定性。
多次實驗結果顯示，嵌入式微型無人機在使用 Simple Lucas-Kanade 光流可
有效進行避障，特別適用於室內環境。此外，模擬結果驗證了 DIS 與 SIF 搭配
Flowdep 在室內外環境的可行性，為未來嵌入式微型無人機開發提供重要參考。

This study investigates monocular vision-based obstacle avoidance methods for
embedded micro aerial vehicles (MAVs), aiming to address the limited computational
efficiency of optical flow algorithms on low-power platforms.
Three optical flow algorithms are evaluated. The Simple Lucas-Kanade method,
with low computational cost, has been implemented on embedded systems but suffers
from limited accuracy. To explore more effective options, the DISand SIF algorithms
are tested in simulation. DIS offers the highest accuracy but requires heavy computation,
while SIF achieves a balance between accuracy and efficiency, making it more suitable
for embedded MAVs.
The study also integrates the Flowdep algorithm in simulation to convert optical
flow into depth information for improved obstacle detection. Two avoidance strategies
are proposed: a simple left-right image partition for embedded systems, and a regionbased segmentation approach in simulation to reduce redundant data and enhance
stability.
Experiments show that Simple Lucas-Kanade enables effective obstacle avoidance
in indoor settings on embedded MAVs. Simulations further validate the feasibility of
using DIS and SIF with Flowdep in both indoor and outdoor environments. These
results provide a foundation for developing efficient embedded vision systems for MAV
navigation.