Considering that scene flow estimation has the capability of the spatial domain to focus but lacks the coherence of the temporal domain, this study proposes long-term scene flow estimation (LSFE), a comprehensive task that can simultaneously capture the fine-grained and long-term 3D motion in an online manner. We introduce SceneTracker, the first LSFE network that adopts an iterative approach to approximate the optimal 3D trajectory. The network dynamically and simultaneously indexes and constructs appearance correlation and depth residual features. Transformers are then employed to explore and utilize long-range connections within and between trajectories. With detailed experiments, SceneTracker shows superior capabilities in addressing 3D spatial occlusion and depth noise interference, highly tailored to the needs of the LSFE task. We build a real-world evaluation dataset, LSFDriving, for the LSFE field and use it in experiments to further demonstrate the advantage of SceneTracker in generalization abilities.

Achieving reliable and safe autonomous driving in off-road environments requires accurate and efficient terrain traversability analysis. However, this task faces several challenges, including the scarcity of large-scale datasets tailored for off-road scenarios, the high cost and potential errors of manual annotation, the stringent real-time requirements of motion planning, and the limited computational power of onboard units. To address these challenges, this paper proposes a novel traversability learning method that leverages self-supervised learning, eliminating the need for manual annotation. We adopt bird's-eye view (BEV) as the input representation to better support ego-centric coordinate representation and integration with motion planning. During vehicle operation, the proposed method conducts online analysis of traversed regions and dynamically updates prototypes to adaptively assess the traversability of the current environment, effectively handling dynamic scene changes. We evaluate our approach against state-of-the-art benchmarks on both public datasets and our own dataset, covering diverse seasons and geographical locations. Experimental results demonstrate that our method achieves superior performance over recent approaches in terms of Area Under the Receiver Operating Characteristic curve (AUROC), Average Precision (AP), and F1-score. Additionally, real-world vehicle experiments show that our method operates at 10 Hz, while a 5.5 km autonomous driving experiment further validates the generated traversability cost map's compatibility with downstream motion planning.

4D millimeter-wave radar provides high-precision ranging capability and exhibits strong robustness under adverse weather and low-visibility conditions, but its point clouds are relatively sparse and suffer from severe elevation-angle measurement noise. Monocular cameras, by contrast, provide rich semantic information and high recall, yet are fundamentally limited by scale ambiguity. To exploit the complementary characteristics of these two sensors, this paper proposes a radar-camera fusion 3D multi-object tracking framework that does not rely on complex 3D annotated data. First, on the radar signal-processing side, a Gaussian distribution-based adaptive angle compression method and IMU-based velocity compensation are introduced to effectively suppress measurement noise, and an improved DBSCAN clustering scheme with recursive cluster splitting and historical static-box guidance is employed to generate high-quality radar detections. Second, a disparity-domain metric depth recovery method is proposed. This method uses filtered radar points as sparse metric anchors, performs robust fitting with RANSAC, and applies Kalman filtering for temporal smoothing, thereby converting the relative depth output of the visual foundation model Depth Anything V2 into metric depth. Finally, a hierarchical fusion strategy is designed at both the detection and tracking levels to achieve stable cross-modal state association. Experimental results on a self-collected dataset show that the proposed method achieves an overall MOTA of 77.93%, outperforming single-modality baselines and other comparison methods by 11 to 31 percentage points. This study provides an effective solution for low-cost and robust environment perception in complex dynamic scenarios.