基於卡爾曼濾波之條件檢測和注意力機制的多物件追蹤演算法

Abstract

隨著自動駕駛技術的蓬勃發展,多物件追蹤的需求日益迫切。本研究提出兩種創新方法以應對複雜道路環境下的挑戰。 首先,我們提出基於關鍵點的多物件追蹤方法。該方法使用無錨點的關鍵點檢測,降低計算資源需求,同時通過改良的深度學習模型架構提取穩健特徵。我們引入背景抑制方法降低誤判,並結合前後幀信息擷取物件運動特徵。針對遮蔽和過曝情況,我們創新性地結合卡爾曼濾波器與檢測器,根據檢測信心動態調整追蹤策略。實驗結果表明,該方法在KITTI追蹤資料集上達到了92.51%的MOTA,優於現有方法,同時保持了約18 FPS的實時性能。 其次,針對密集複雜場景,我們提出基於注意力機制的追蹤方法。該方法結合transformer架構的DINO檢測器和多頭注意力機制,有效捕捉長期物件關聯。我們還引入重識別機制,增強長時間遮蔽後的追蹤能力。在MOT17數據集上,該方法達到74.8%的MOTA,特別適合處理複雜密集場景。 兩種方法各具優勢,為不同場景下的多物件追蹤提供了有效解決方案。前者在計算效率和通用性方面表現優異,後者則在處理複雜場景和長期依賴關係方面更具優勢。

With the rapid development of autonomous driving technology, the demand for efficient multiple object tracking has become increasingly urgent. This study proposes two innovative methods to address the challenges in complex road environments. First, we introduce a keypoint-based multiple object tracking method. This approach utilizes anchor-free keypoint detection to reduce computational resources while extracting robust features through an improved deep learning model architecture. We implement a background suppression technique to minimize false detections and incorporate information from adjacent frames to capture object motion characteristics. To address occlusion and overexposure scenarios, we innovatively combine a Kalman filter with the detector, dynamically adjusting the tracking strategy based on detection confidence. Experimental results demonstrate that this method achieves a MOTA of 92.51% on the KITTI tracking dataset, outperforming existing methods while maintaining real-time performance at approximately 18 FPS. Second, targeting dense and complex scenarios, we propose an attention-based tracking method. This approach integrates a DINO detector with transformer architecture and a multi-head attention mechanism, effectively capturing long-term object associations. We also incorporate a re-identification mechanism to enhance tracking capabilities after prolonged occlusions. On the MOT17 dataset, this method achieves a MOTA of 74.8%, particularly excelling in handling complex, dense scenarios. Both methods offer unique advantages, providing effective solutions for multiple object tracking in various scenarios. The former excels in computational efficiency and versatility, while the latter demonstrates superior performance in handling complex scenes and long-term dependencies. Future research will explore combining the strengths of both methods to further enhance multiple object tracking performance.