針對服務型自主移動機器人在實際場域中對低延遲、高精度自主導航與即時障礙迴避的需求， 本論文設計並實現一套基於邊緣運算的物件偵測與語義地圖構建系統。系統採用經結構化剪枝與後訓練量化的YOLOv5s 模型，僅在筆電端即可穩定達成高幀率的物件偵測，以確保能即時生成障礙物資訊並驅動導航決策。相較於傳統依賴雲端計算的做法，本論文透過本地化推論大幅降低網路延遲與頻寬需求，並提升系統的可靠性與資料隱私保護。
先將相機偵測框映射至2D LiDAR 深度點雲， 透過幾何校正與時間同步構建初步的語義地圖。進一步提升地圖精度與定位穩定性，本論文引入擴增卡爾曼濾波機制， 將LiDAR、里程計及相機視覺量測做多感測融合， 抑制感測雜訊及漂移。系統在長廊與開放空間等室內場域進行實驗評估，用以觀測指標包括地標映射精度、導航成功率及推論延遲等成效。結果顯示採用擴增卡爾曼濾波前後相比，能顯著降低地標映射誤差並提高導航成功率；更顯著地是系統在室內場域中均能保持長時間穩定運行， 能即時偵測並避開動態及靜態障礙物， 而延遲及資源占用均限定在可接受範圍。

This thesis illustrates the needs for low-latency, high-precision autonomous navigation and real-time obstacle avoidance in service-oriented mobile robots operating in real-world environments by designing and implementing an edge-computing–based perception system for object detection and semantic map construction. The proposed system employs a structurally pruned and post-training quantized YOLOv5s model, which runs entirely on a medium-speed laptop computer to deliver high frame-rate object detection and ensure timely generation of obstacle information for navigation decisions. Different from traditional cloud-driven approaches, our edge-localized inference markedly reduces network latency and bandwidth
usage while enhancing system reliability and data privacy.
First, detected bounding boxes from the camera are projected into 2D LiDAR point clouds, and through geometric calibration and time synchronization. An initial semantic map is generated. To further improve map accuracy and localization stability, Extended Kalman Filters can fuse measurements from LiDAR, odometry, and visual sensing, and effectively suppressing noises and drifts. Practical validations conducted in building corridor and open-space scenarios prove the key metrics among landmark mapping accuracy, navigation success rate, and inference latency. It is evident that the Extended Kalman Filter substantially reduces mapping errors and increases navigation successfulness, whereas the system stability, real-time detection of both dynamic and static obstacles, and acceptable latency and resource usage throughout operation are all ballanced.