多物件追蹤技術應用相當廣泛，不論是從監視器畫面進行車流監控、人流監控、行人追蹤、球場上球員走位的戰術分析都會用到多物件追蹤的技術，其主要任務是將一段影片中分布在不同幀的偵測框正確的關聯起來，困難的地方在於當目標被長時間遮蔽、消失、場景複雜的情況下容易發生追蹤錯誤的情況，雖然已有許多研究提出不同的追蹤策略來解決此問題，但追蹤結果仍有可進步的空間。 為了提高多物件追蹤的準確度，本論文基於ByteTrack架構上提出一個兩階段的Online多物件追蹤方法: MEDIATrack，我們將Kalman Filter更換為NSA Kalman Filter、引入外觀特徵作為資料關聯參考資訊，並設計懲罰機制去緩解在場景複雜所出現的錯誤情況，此外也移除歷史軌跡中未激活軌跡的機制，直接將高信心值未匹配上的偵測框新增至歷史軌跡，使得本研究在MOT17資料集達到79.3(%) MOTA，達到了state-of-the-art的水準。 ;Multi-object tracking (MOT) is widely applied to traffic flow monitoring, human flow monitoring, pedestrian tracking, or tactical analysis of players on the courts. It associates the detection boxes with tracklets for each frame in the video. The challenges of MOT include long-term occlusions, missing detections, and complex scenes. Although many trackers have proposed to solve these problems, the tracking results still have room for improvement. In this thesis, we propose a solution named MEDIATrack, which is a two-stage online multi-object tracking method based on the ByteTrack. We replace the Kalman Filter with the NSA Kalman Filter, introduce appearance features for track association, and design a punishment mechanism to alleviate errors in complex scenes. In addition, we remove the nonactivated strategy, and the high-score unmatched detection boxes are directly added to the tracklets. On MOT17, we achieve 79.3 MOT Accuracy and state-of-the-art performance.
