非監督式的三維點雲抗遮蔽多物體追蹤

Abstract

自駕系統通過感知儀器，例如光達，獲取外界資訊，並用其進行路徑規劃與決策。在路徑規劃中，準確的三維多物體追蹤至關重要，因為它有助於提高自駕系統的安全性與效率。然而，光的直線特性限制了感知資訊的搜集，常常導致遮蔽，從而嚴重影響物體追蹤的品質。遮蔽還會引發物件識別碼(ID)切換，導致歷史資訊的丟失，不利於分析車輛的駕駛模式或路徑。本研究提出了一種非監督式的抗遮蔽追蹤演算法。在資料關聯階段，使用物體偵測位置和點雲的外觀特徵，設計多階段的資料關聯，來彌補遮蔽期間的動作預測誤差。點雲的外觀特徵由正態分布變換(Normal Distributions Transform)提取，基於機率統計的規則導向設計，實現非監督的特徵提取與相似度比較。此外，本研究改進了現有的物件追蹤演算法，使其能兼顧輸出品質與物體的最大允許遮蔽時間，可定義為追蹤強度。本研究在路口(Wayside)和車載(KITTI)兩個資料集上進行驗證。在Wayside資料集上，使用模擬遮蔽情形的物體偵測結果(GTdet)進行追蹤，能達到95.81%的準確度，且沒有產生物件識別碼切換，整體表現遠勝於傳統的AB3DMOT。在KITTI驗證集上，和CONF_MOT比較，準確度從69.09%提升至71.11%，且物件識別碼切換次數從8下降至4。本研究提出的方法能有效改善遮蔽問題，且在靜止或移動光達所採集的點雲數據上均具有良好的通用性。

Occluded images or point cloud data can lead wrong detection on object class and sizes. This work proposes an unsupervised occlusion resilience tracking algorithm for point cloud data. In the data association phase, the object detection information and point cloud features are used to compensate motion prediction errors caused by occlusion. The features of the point cloud are extracted using Normal Distributions Transformation, which is designed based on probabilistic statistical rules to achieve unsupervised point cloud feature extraction and similarity comparison. Additionally, this work improves the existing tracking algorithms, to balance output quality with the maximum allowable occlusion time of the object, namely tracking robustness. The proposed algorithm is evaluated on two different datasets: Wayside and KITTI. On Wayside dataset, using simulated detection results under occlusion scenarios for tracking achieves an MOTA of 95.81% without any ID switches, significantly outperforming traditional AB3DMOT. On KITTI validation set, compared to CONF_MOT, the MOTA improves from 69.09% to 71.11%, and the number of ID switches decreases from 8 to 4. The experiment results demonstrate that the proposed method effectively mitigates occlusion issues and demonstrates good generality for point cloud collected by both stationary and mobile LiDAR.