深度增強追蹤器減輕多目標遮擋和同質外觀問題於室內追蹤

Abstract

多目標追蹤長期以來一直是人們感興趣的議題，因為它在許多計算機視覺應用中發揮著重要作用。現有研究多為戶外追蹤設計，如影像監控和自動駕駛。然而，戶外追蹤場景中物體的行為不能完全反映室內追蹤環境中的挑戰。在戶外追蹤場景中，行人和車輛通常在一條簡單的直線路徑上均勻地從一個地方移動到另一個地方，且行人的外觀通常是差異巨大的。相比之下，在室內場景中，例如舞蹈編排表演，舞者的動態行為導致嚴重的遮擋，類似的表演服裝呈現出同質的外觀問題。室內追蹤中的這些嚴重遮擋和同質外觀問題導致現有追蹤器的性能明顯下降。在本文中，我們提出了一個深度增強的多目標追蹤框架和語義匹配策略與場景感知相結合親和力測量方法可顯著減輕遮擋和同質外觀的問題。此外，我們引入了室內追蹤數據集並增加了現有基準數據集的多樣性用於室內追蹤評估。我們設計實驗去評估我們的追蹤器和現有的追蹤器，在我們提出的室內追蹤數據集和最新的 MOT17 和 MOT20 測試數據集上，我們的方法始終如一在令人信服的 HOTA 指標上優於其他追蹤器。相較於實驗中第二好的追蹤器 DeepSORT 相比，我們提出的追蹤器大大降低身份轉換的數量將近 20\% 在我們提出的室內追蹤數據集中。

Multiple-object tracking has long been a topic of interest since it plays an important role in many computer vision applications. Existing works are mostly designed for outdoor tracking, such as video surveillance and autonomous driving. However, the behaviors of objects in outdoor tracking scenarios do not fully reflect the tracking challenges in indoor tracking environments. In outdoor tracking scenarios, pedestrians and vehicles usually move uniformly from place to place on a simple straight path, and target appearances are usually different. In contrast, in indoor scenarios, such as choreographed performances, the dynamic behaviors of dancers lead to severe occlusions, and similar costumes present a homogeneous appearance problem. These severe occlusion and homogeneous appearance problems in indoor tracking lead to noticeable degradation in the performance of existing works. In this paper, we propose a depth-enhanced tracking-by-detection framework and a semantic matching strategy combined with a scene-aware affinity measurement method to mitigate occlusion and homogeneous appearance problems significantly. In addition, we introduce an indoor tracking dataset and increase the diversity of existing benchmark datasets for indoor tracking evaluation. We conduct experiments on both the proposed indoor tracking dataset and the latest MOT benchmarks, MOT17 and MOT20. The experimental results show that our method consistently outperforms other works on the convincing HOTA metric across the benchmarks and greatly reduces the number of identity switches by 20% compared to that of the second-best tracker, DeepSORT, in our proposed indoor MOT benchmark dataset.