3D SOT, a core task in computer vision and autonomous driving, aims to continuously localize target objects across consecutive point cloud frames using their initial states. The proliferation of 3D sensing technologies (e.g., LiDAR) has established point clouds as a critical data source for object tracking, primarily due to their illumination invariance and inherent geometric information. However, inherent challenges including sparsity, spatial disorder, and non-uniform density distribution fundamentally complicate point cloud processing. Furthermore, object deformation during motion, dynamic occlusion, and interference from inter-class similarity present ongoing tracking challenges.

In practical scenarios, sparsity manifests acutely; for instance, long-distance targets (e.g., pedestrians 100 m away) may contain only 20–50 points, making it difficult to extract stable features; in rainy or foggy weather, LiDAR point cloud density can drop by 60%, blurring target contours. Spatial disorder, referring to the fact that point clouds have no fixed order, renders traditional 2D feature extraction methods—dependent on grid structures—inapplicable, necessitating specialized architectures compatible with disorder. Non-uniform density distribution further causes near-dense and far-sparse biases in feature extraction, affecting global localization accuracy. Furthermore, object deformation during motion, dynamic occlusion, and interference from inter-class similarity present ongoing tracking challenges—deformation leads to mismatches between templates and current frames; occlusion may cause target point clouds to disappear entirely; and inter-class similarity easily triggers mistracking in complex traffic flows.

Early approaches predominantly adapted 2D visual similarity matching paradigms, determining object positions through template-search region correspondences [1,2]. However, 2D methods rely heavily on texture information, which is scarce in point clouds, leading to inherent mismatches with 3D point cloud characteristics. For example, 2D cross-correlation operations fail to handle point cloud disorder, resulting in unstable matching in sparse scenarios. Although conceptually straightforward, these methods were fundamentally limited by point cloud sparsity, exhibiting significant difficulty in discriminating geometrically similar distractors. Moreover, the reliance on handcrafted or computationally intensive bounding box estimation techniques—such as Kalman filtering [3]—introduces significant computational overhead. Traditional Kalman filtering-based methods often require 𝒪(n3) matrix operations per frame, which poses a major bottleneck for real-time deployment in resource-constrained environments.

With the rapid growth of point cloud data and advancements in computational resources, end-to-end learning frameworks have made significant strides by introducing collaborative voting schemes for object center candidate generation [4]. For instance, VoteNet [5] aggregates sparse points into object centers through point-wise voting, effectively addressing the sparse-to-dense localization bottleneck. This approach notably improves both localization accuracy and computational efficiency, thereby reinforcing the prevalence of matching paradigms. Subsequent refinements have further enhanced feature matching via employing voxel-based representations [6], and augmenting local geometric features [7]. Despite these advances, fundamental challenges remain—particularly in mitigating tracking drift caused by inherent point cloud sparsity. Even state-of-the-art methods exhibit a 30% higher drift rate in extremely sparse scenarios (≤30 points) compared to dense scenarios.

To overcome the constraints of matching paradigms, motion-centric modeling has emerged as a promising alternative. These techniques explicitly model inter-frame object motion dynamics, shifting focus from appearance similarity to motion patterns—often through techniques such as pose transformation estimation via foreground point segmentation [8]. This methodology exhibits superior robustness under occlusion and point cloud sparsity conditions. However, dependence on local motion cues from neighboring frames fundamentally limits long-term motion pattern capture. Moreover, segmentation inaccuracies propagate pose estimation errors, leading to cumulative drift. Recent breakthroughs have begun to mitigate these issues by integrating multimodal data fusion [9], temporal context modeling [10], and Transformer architectures [11], which significantly enhance feature extraction and correspondence accuracy. Simultaneously, progress in point cloud representation learning and lightweight network design enables novel accuracy-efficiency tradeoffs [12].

To establish a rigorous research boundary, we adopt the formal problem definition proposed by Hodaň et al. [13]. Mathematically, the 3D SOT task aims to estimate the optimal bounding box Bt of a target in the current frame Pt, given the initial state B0 and a sequence of historical observations 𝒫0:t−1, by maximizing the posterior probability P(Bt|Bt−1,Pt). Within this theoretical framework, this review explicitly restricts its scope to ‘complex dynamic scenarios’ in autonomous driving, which are characterized by high sparsity (objects with fewer than 50 points), severe occlusion, and non-rigid deformation. Consequently, we focus exclusively on tracking algorithms designed for unstructured and sparse LiDAR point clouds, extending the discussion from foundational Siamese-based frameworks to emerging Motion-Centric and SSM approaches.

Despite the rapid growth of this field, a systematic survey dedicated to 3D SOT is notably absent. Existing reviews primarily focus on 3D object detection or 2D tracking, often overlooking the specific challenges of temporal coherence in sparse point clouds. Previous surveys on multi-object tracking also differ significantly in problem formulation. Therefore, a comprehensive analysis of 3D SOT algorithms is urgently needed.

To ensure a rigorous and objective consolidation, we implemented a systematic literature search and filtering methodology. The retrieval was conducted across major academic databases, including IEEE Xplore, Web of Science, Google Scholar, and the Computer Vision Foundation Open Access, covering the period from January 2019 to January 2026. The search strategy utilized Boolean logic with keyword combinations such as “Point Cloud,” “LiDAR,” “3D,” “Single Object Tracking,” and “Deep Learning.” This initial phase yielded over 300 candidate entries.

To isolate the most significant contributions, we applied a strict multi-stage screening process. First, the scope was restricted exclusively to 3D SOT, excluding studies focused solely on 2D tracking or Multi-Object Tracking unless they offered specific contributions to the SOT domain. Second, priority was given to peer-reviewed articles published in high-impact journals (e.g., IEEE TPAMI, TITS) and top-tier conferences (e.g., CVPR, ICCV, ECCV, AAAI), with a requirement that the methods be validated on standard benchmarks. Finally, to capture the rapid evolution of emerging architectures like Transformers and SSMs, we emphasized timeliness, with approximately 85% of the selected studies published between 2021 and 2026. Based on these criteria, 77 key publications were ultimately selected for in-depth analysis in this review.

This survey provides a comprehensive and systematic analysis of the technological evolution in point cloud object tracking. Over the past few years, the field has transitioned from early appearance-based models to more advanced architectures that incorporate motion reasoning and spatio-temporal context modeling. To ground the discussion, a representative pipeline commonly adopted in modern 3D tracking systems is illustrated in Fig. 1. It highlights the classical Siamese network structure, which forms the basis for many current methods, along with its typical extensions in feature extraction, similarity matching, and template updating.

Corresponding to the general Siamese-based tracking framework, the structure of this paper is organized as follows: (1) Backbone Networks: This section reviews the development of feature extraction architectures, from early hand-engineered frameworks and voxel-based methods to modern Transformer and Mamba-based models, analyzing their capabilities in capturing geometric structures and spatio-temporal relationships. (2) Decision Mechanisms: We categorize and compare three core paradigms—similarity matching, motion modeling, and attention mechanisms—elucidating their respective strengths in handling appearance variations, motion dynamics, and long-range dependencies. (3) Model Update Strategies: This part focuses on dynamic template optimization, including shape compensation, temporal memory management, and adaptive parameter adjustment, exploring how these strategies mitigate tracking drift under occlusion and sparsity. (4) Experiments: A detailed analysis of mainstream metrics and datasets (KITTI [14], NuScenes [15]) is provided to establish a standardized performance comparison framework. (5) Discussion: This section synthesizes the technological evolution of tracking paradigms and analyzes persistent bottlenecks. (6) Conclusion: This section summarizes the systematic review of 3D single object tracking and outlines future directions.

To the best of our knowledge, this is the first comprehensive survey dedicated to point cloud object tracking. Its key contributions are fourfold: (1) Systematic Taxonomy: We implement a unified classification for tracking paradigms, clarifying the evolutionary logic from 2D adaptation to 3D-specific innovations. As this is a review paper, we do not propose any novel framework. (2) In-Depth Technical Analysis: By dissecting core modules (feature extraction, decision-making, updates), we reveal the theoretical limitations of each method and their performance bottlenecks under real-world constraints. (3) Benchmark Synthesis: A cross-dataset comparison of the state-of-the-art methods is conducted to quantify the effectiveness of different strategies under varying scenarios. (4) Research Outlook: By synthesizing theoretical advances and empirical results, this survey aims to serve as a foundational reference for researchers and engineers in the field, facilitating the development of next-generation point cloud tracking technologies. It is important to clarify that this paper serves as a comprehensive taxonomy and performance analysis of existing methods. While we systematically synthesize the evolution of the field, this work does not propose a novel tracking algorithm but rather consolidates recent years of developments to guide future research expectations.

Early approaches addressed the unstructured nature of point clouds by leveraging voxelization, a process that maps 3D points to regular grid spaces [17]. This enabled the adaptation of 2D CNN architectures for spatio-temporal motion modeling, offering a pragmatic solution to data disorder [18,19]. Nevertheless, voxelization inherently introduced quantization artifacts, degrading geometric fidelity and compromising the representation of fine-grained motion trajectories—limitations that motivated subsequent innovations.

The transition to direct point processing represented a paradigm shift in computational efficiency. PointNet [20] pioneered this by using shared multi-layer perceptron (MLP) for permutation invariance, though it failed to capture local structures. PointNet++ [21] addressed this by introducing a hierarchical framework with sampling and grouping, establishing a standard for local feature extraction. However, its recursive sampling creates computational bottlenecks.

Subsequent works focused on refining local feature aggregation. PointCNN [22] introduced X-transformation to enable convolution on unordered points. To enhance geometric reasoning, Relation-Shape CNN [23] and PointWeb [24] modeled inter-point relations and contextual interactions. FoldingNet [25] extended this line of work by introducing folding operations to generate point clouds from latent spaces, achieving state-of-the-art (SOTA) performance in shape completion tasks. Meanwhile, PointConv [26] and PAConv [27] improved computational efficiency by learning dynamic weight functions for irregular point densities.

In summary, point-based backbones process raw coordinates directly, preserving fine-grained geometric details essential for precise localization. However, their reliance on iterative sampling and grouping often creates a computational bottleneck.

Graph neural networks naturally align with point cloud topology for modeling point cloud topology, given their capacity to capture pairwise relationships. DGCNN [28] pioneered this direction with dynamic graph convolution, constructing k-NN graphs in feature space and learning edge representations via EdgeConv [29]. While this significantly enhanced local geometric modeling, dynamic graph updates resulted in quadratic computational complexity as point counts increased. Subsequent works mitigated this issue: SpiderCNN [30] designed adaptive neighborhood expansion strategies, and Point-GNN [31] refined graph-based message passing to better capture structural dependencies.

Attention-based architectures revolutionized point cloud processing by enabling adaptive feature weighting. Point-BERT [32] introduced a novel pretraining paradigm through Masked Point Modeling.

Transformers [33], with their capacity to model global context, were adapted to point clouds by PCT [34]—the first work to integrate standard Transformer architectures. To handle spatial unorderedness, PCT [34] introduced coordinate embedding (injecting 3D positional information via encoding) and offset attention (dynamically adjusting geometric offsets in attention calculations), thereby strengthening the perception of spatial relationships. Swin3D [35] optimized Transformer efficiency for 3D scenes by adopting window-based attention (inspired by 2D vision): it partitions point clouds into spatial windows, computes self-attention within windows, and uses window shifting for cross-window interaction, balancing local geometric accuracy and computational cost. Point Transformer [36] further reduced complexity by restricting attention to overlapping windows, thereby striking a finer balance between performance and efficiency.

Sequence-based methods transformed unordered point clouds into ordered structures, thereby facilitating spatio-temporal modeling. Point2Sequence [37] pioneered this paradigm by converting points into sequences to enhance the modeling of dynamic scene continuity.

Recent advances have leveraged state-space models: PointMamba [38] introduced Mamba architectures to point cloud processing by using Z-order curve partitioning to map unordered points to spatially localized sequences. Combined with bidirectional selective scanning, it maintained spatio-temporal modeling capabilities while reducing GPU memory consumption by over 60%. Its MAE-style pretraining further enhanced the representation of geometric structures. SMamba [39] optimized this framework for sparse point clouds by employing adaptive sparsity-aware mechanisms to boost the efficiency of long-range feature modeling. Additionally, submanifold sparse convolutions mitigated the “near-dense, far-sparse” bias, which is critical for long-range tracking tasks.

In the field of 3D object tracking, decision mechanisms based on Similarity Matching (often referred to in the literature as Siamese-based tracking or Appearance-based tracking) have long occupied a central position. Fundamentally, these methods achieve localization by modeling the feature similarity between the template object and the search region. Such approaches typically adopt a Siamese network architecture: weight-sharing feature extraction branches process the template point cloud and search region point cloud, respectively, with efficient similarity calculation modules designed to achieve object matching [40].

As a foundational work in this direction, SC3D [41] was the first to introduce Siamese networks into the field of point cloud tracking. The method generates candidate object proposals via Kalman filtering [3], uses an encoder to map the template and candidate point clouds to a compact latent space, and then performs object matching via cosine similarity. However, its mechanism relying on iterative optimization for proposal generation not only rendered end-to-end training challenging but also introduced significant bottlenecks in computational efficiency.

In contrast to appearance-based matching strategies, motion modeling effectively mitigates the adverse effects of point cloud sparsity, occlusion, and sensor noise on tracking robustness by explicitly characterizing the motion dynamics of objects in spatio-temporal sequences, thus exhibiting unique advantages in complex dynamic scenarios.

The introduction of Transformer architectures [33] marked a pivotal shift in feature modeling paradigms. By virtue of the self-attention mechanism’s capability to efficiently model long-range dependencies, Transformer architectures broke through the limitations of traditional Siamese networks that rely on fixed matching modules, thereby exhibiting superior feature fusion and object localization capabilities in complex dynamic scenarios.

Note that the inter-frame attention discussed here differs from the intra-frame attention in Section 2.4. While Section 2.4 focused on intra-frame attention for extracting local geometric features within the backbone, this section addresses inter-frame attention as a decision mechanism to model the global correlation between the template and the search region.

PTTR [11], as a pioneering work in this field, first constructed a point-relation Transformer framework. It strengthened local feature interaction between the template and search region via self-attention mechanisms, and achieved global-scale feature matching and alignment through cross-attention mechanisms. This end-to-end decision paradigm abandoned the traditional approach of manually designed similarity metrics, enabling the model to adaptively learn geometric correlation features of inter-frame point clouds. PTTR++ [55] further optimized the architecture design by proposing an innovative Point-BEV dual-branch Transformer fusion strategy. This method extracts fine geometric features via MLPs in the point cloud domain, while generating dense feature maps through voxelization and max pooling in the BEV domain. It achieves complementary enhancement of point cloud details and BEV global structures via cross-branch attention mechanisms, thereby realizing multi-scale feature integration. CXTrack [56] introduced an object-centric Transformer, which propagates contextual information and object cues from historical frames to the current frame through attention, effectively mitigating feature degradation caused by object occlusion. SyncTrack [57] completely abandoned the Siamese structure and proposed a single-branch synchronous architecture. By concatenating template and search region point clouds and feeding them into the Transformer, it naturally integrates feature extraction and matching processes using a dynamic attention matrix, significantly improving accuracy while maintaining real-time performance (45 FPS).

To address the feature sampling challenge caused by point cloud sparsity, PTTR [11] proposed a relationship-aware sampling strategy. By dynamically selecting highly correlated points based on attention response values between the template and search region, it significantly reduced geometric information loss induced by random sampling. SyncTrack [57] further built upon this optimization by designing an attention-point sampling Transformer, which guides key point selection via attention weights. This approach preserves more discriminative features while maintaining computational efficiency.

Through continuous innovation in Transformer architectures, these methods have driven a paradigm shift in 3D object tracking from manual feature matching to adaptive spatiotemporal modeling. Their technological evolution not only reflects exploratory efforts in multi-scale feature fusion but also demonstrates the ongoing optimization of the balance between real-time performance and tracking accuracy.

To address key challenges such as point cloud sparsity and object occlusion, dynamic updates of tracking models are achieved through geometric structure repair and feature representation optimization. The technical core of such methods lies in addressing challenges related to object representation under sparse point cloud scenarios through approaches including point cloud geometric completion, dynamic template optimization, and virtual feature generation. Their core objective is to maintain the template’s capability to precisely characterize the target’s geometric properties and appearance features throughout the tracking process.

Temporal memory and model update strategies have provided key technical pathways for addressing challenges such as object occlusion, deformation, and template drift in long-term tracking by exploiting spatiotemporal correlation information from historical frames. These methods rely on constructing efficient historical information management mechanisms, which enable adaptive updates of tracking models to accommodate object motion patterns through dynamic selection and fusion of historical frame features or templates. Their technological evolution has been prominently reflected in the paradigm shift from independent single-frame processing to temporal dependency modeling.

Addressing the issue of trajectory interruption caused by object occlusion and point cloud sparsity in complex dynamic scenarios, recent studies have significantly improved the recovery capability of tracking systems under extreme conditions by constructing precise failure detection metrics and efficient relocalization strategies.

DeepPCT [70] systematically constructed a detection-relocalization closed-loop mechanism. The core innovation of this method lies in proposing the center intersection-over-union (CIoU) metric, which extends the traditional IoU by introducing a normalized center distance. This effectively addresses the challenge of quantifying spatial relationships for non-overlapping bounding boxes. Compared to traditional metrics that only focus on region overlap, CIoU can more accurately characterize the relative positional deviation between predicted and ground truth boxes, making it particularly suitable for object localization evaluation in sparse point cloud scenarios. Upon detecting that the CIoU value falls below a predefined confidence threshold, the system automatically activates the relocalization module. This module invokes a pre-trained 3D object detector to perform high-density point cloud reconstruction and object proposal generation in the global search region, thereby achieving geometric feature matching and position correction for drifted targets. Experimental data shows that this mechanism increases the success rate by 18 percentage points in long-term tracking tasks on the KITTI [14] dataset. Notably, in extremely sparse scenarios where the number of single-object points is less than 150, the relocalization response time is improved by 40% compared to baseline methods, significantly reducing the risk of tracking failure caused by feature loss.

In the field of implicit anti-degradation research, LTTR [71] constructed a Transformer-based encoder-decoder framework, enhancing feature robustness in sparse scenarios by explicitly modeling point cloud region relationships. This method breaks through the local neighborhood limitation of traditional point cloud processing, capturing long-range dependencies between any two points using a multi-head self-attention mechanism while retaining the prior knowledge of spatial distribution of point clouds through a positional encoding module. STNet [72] and CorpNet [73] represent technical pathways for enhancing robustness via network structure optimization. STNet [72] designed an iterative cross-self-feature enhancement module: during the feature interaction phase, it embeds the global geometric information of the template object into the current frame’s feature space via cross-attention, then aggregates locally semantically similar points using a self-attention mechanism, forming a multi-round progressive feature optimization process. This mechanism effectively suppresses error accumulation in short-term object occlusion scenarios, maintaining tracking trajectory continuity through iterative calibration of historical frame features. CorpNet [73] focused on multi-scale feature encoding, constructing a correlation pyramid structure with 512/256/128-point hierarchies. It simultaneously preserves high-level semantic information and low-level geometric details during feature extraction. Through a cross-layer feature fusion strategy, this method significantly enhances structural awareness capability in sparse point cloud regions, reducing the probability of tracking failure caused by local feature loss by 23%.

To ensure a comprehensive and rigorous evaluation of 3D single object tracking methods, we conducted systematic comparisons on two benchmark datasets: KITTI [14] and NuScenes [15], which are widely recognized for their representativeness in autonomous driving scenarios. Specifically, the KITTI dataset comprises 21 video sequences for training and 29 for testing, with each sequence containing synchronized LiDAR point clouds and camera images captured in urban, rural, and highway environments. Due to the inaccessibility of official test set labels, we further split the training set into three non-overlapping subsets (train/val/test) following standard protocols, ensuring unbiased performance validation. For NuScenes, a larger-scale dataset with richer scene diversity, it includes 1000 scenes covering various weather conditions (e.g., sunny, rainy, foggy) and time periods (day and night), which are partitioned into 700 training scenes, 150 validation scenes, and 150 test scenes to support comprehensive model generalization analysis.

In terms of evaluation categories, we focused on five typical object classes critical to autonomous driving: Car, Pedestrian, Truck, Trailer, and Bus. These categories were selected based on their high frequency in traffic scenarios and the distinct challenges they pose (e.g., small size for Pedestrians, large volume for Trucks/Trailers), ensuring the evaluation covers diverse tracking scenarios.

The quantitative evaluation results of state-of-the-art algorithms on the KITTI and NuScenes datasets are presented in Tables 1 and 2, respectively (“-” indicates that there are no relevant experiments for this method; the best and second-best results are highlighted in red and blue, respectively). To ensure a fair comparison, we adopt Success and Precision defined in one-pass evaluation (OPE) [75,76] as the evaluation metrics. Success denotes the Area Under the Curve (AUC) for the plot of the ratio of frames in which the IoU between the predicted box and the ground truth box exceeds a threshold (ranging from 0 to 1). Meanwhile, Precision denotes the AUC for the plot of the ratio of frames in which the distance between their centers (i.e., the predicted box and the ground truth box) is within a threshold (ranging from 0 to 2 m).

Table 1 summarizes SOTA 3D single-object tracking performance on the KITTI dataset, with evaluations based on Success and Precision. SiamMo [54] establishes a new SOTA with a mean Success of 72.3 and a mean Precision of 90.1, outperforming MTM-Tracker [50] (70.9/88.4) by 1.4 points. This performance gain stems from its spatio-temporal feature aggregation (STFA) module, which explicitly models cross-frame motion trajectories, which is a critical capability for mitigating partial occlusion and fast motion in urban and highway scenarios.

Table 2 presents performance evaluations on NuScenes—a larger-scale dataset featuring diverse weather conditions (e.g., rain, fog) and sparse long-range point clouds (>50 m), which amplifies generalization challenges. MVCTrack [59] establishes itself as the SOTA (61.20 in mean Success, 73.22 in mean Precision), outperforming SiamMo [54] (60.31/72.68) and P2P [53] (59.84/72.13). Its multi-view cross-modal fusion mechanism integrates RGB texture and LiDAR geometric features, compensating for sparse or noisy points—a critical capability for addressing NuScenes’ variability.

While Tables 1 and 2 provide exhaustive details, it is beneficial to summarize the performance ceilings of different technical routes. Table 3 presents a concise comparison of the representative SOTA methods from each category. As observed, Sequence & Update methods (e.g., M3SOT [69], MVCTrack [59]) currently achieve the highest performance on both datasets, validating the importance of temporal context in handling occlusion. Motion Modeling methods (e.g., P2P [53]) show exceptional robustness on the NuScenes dataset, narrowing the gap with sequence-based methods, whereas traditional Similarity Matching approaches generally lag behind in complex sparse scenarios.

Regarding backbone impact, Transformer-based backbones consistently outperform (e.g., PTTR [11]) consistently outperform PointNet++ [20] based ones by approximately 5% in Success, validating the importance of global context capture in sparse point clouds. In terms of decision mechanism effectiveness, Motion-centric methods (e.g., M2-Track [8]) show superior robustness on rigid objects like cars compared to Similarity Matching, as they explicitly utilize kinematic constraints. However, for non-rigid objects such as pedestrians, attention-based methods show better adaptability. Finally, concerning update strategies, methods utilizing Temporal Memory (e.g., MBPTrack [65]) demonstrate a significant performance boost in long-term tracking scenarios compared to single-frame baselines.

We select BAT [43] as the baseline for visualization because it represents the foundational “Box-Aware” paradigm and serves as a standard open-source benchmark for validating recent advancements.

To mitigate the paucity of visual evaluations in existing 3D point cloud tracking literature, we integrate visualization data from M2-Track [8] and compare it with BAT [43], as illustrated in Figs. 5 and 6. These visual comparisons offer mechanistic insights into the representative tracking sequence, thereby facilitating a nuanced understanding of the efficacy of motion modeling and trajectory estimation in 3D object tracking.

In Scene 0020 (Frame 55), the notable position error of M2-Track (0.077 m vs. BAT’s 0.870 m in single-frame localization) exposes limitations in its motion modeling under fine-grained spatial constraints. The discrepancy between M2-Track’s bounding box and the ground truth across multi-view projections (e.g., XY, XZ views) indicates that its motion-driven localization strategy grapples with precise spatial alignment in dense point cloud clusters. Conversely, BAT exhibits superior single-frame localization accuracy (0.870 m error) and trajectory consistency, attributed to its effective fusion of geometric features and motion priors—underscoring the pivotal role of multi-modal information integration in precise tracking. Notably, BAT’s consistent alignment with the ground truth in both single-frame and trajectory analyses corroborates that fine-grained feature matching, coupled with robust motion regularization, constitutes an effective approach to enhancing tracking precision and temporal stability. This design aligns with methodologies emphasizing spatiotemporal coherence and multi-view geometric consistency, collectively indicating an emerging trend of integrating precise spatial alignment with temporal continuity in 3D tracking.

In the trajectory analysis of Scene 0020, M2-Track’s substantial mean position error (8.086 m) and trajectory divergence highlight the inadequacy of its motion-centric modeling in sustaining long-term tracking consistency under complex spatial dynamics. This shortcoming stands in contrast to BAT’s stable trajectory (mean error 0.460 m), which benefits from adaptive motion correction and multi-frame feature aggregation—mechanisms that bolster robustness by explicitly modeling spatiotemporal dependencies. Furthermore, BAT’s precision in both single-frame localization and trajectory estimation aligns with a convergent technical trajectory of methodologies emphasizing the tight integration of spatial geometry and temporal motion, collectively propelling tracking models toward spatiotemporal co-optimization.

These visual findings not only validate the efficacy of quantitative metrics but also unveil two core trends in contemporary 3D tracking: a paradigm shift from the pursuit of single-frame precision to the assurance of long-term trajectory consistency, and a methodological evolution from motion-only or feature-only modeling to integrated spatiotemporal-feature fusion. As paradigmatic examples, BAT’s multi-view alignment mechanism and M2-Track’s motion modeling strategy demonstrate that precise spatial alignment and robust motion regularization are pivotal strategies for addressing long-term trajectory divergence and fine-grained localization errors in 3D tracking.

The field has witnessed a paradigm shift from 2D-inspired Similarity Matching to 3D Motion Modeling, and recently to Sequence-based learning. This evolution reflects a growing capability to handle the intrinsic disorder and sparsity of point clouds, moving from simple texture-less matching to sophisticated dynamic state estimation.

3D SOT has witnessed substantial advancements, yet critical challenges persist, limiting its robustness in real-world scenarios. Extreme point cloud sparsity remains a primary bottleneck: long-range targets (e.g., pedestrians 100 m away with 20–50 points) or sensor noise in adverse weather (60% density reduction in rain/fog) degrade feature stability, leading to a 30% higher drift rate in state-of-the-art methods compared to dense scenarios. This sparsity exacerbates mismatches in similarity matching frameworks and propagates errors in motion modeling pipelines.

Long-term occlusion further disrupts tracking continuity. Motion-centric methods relying on local inter-frame cues fail to capture non-rigid dynamics or complex trajectories, while similarity-based approaches lose discriminative features, resulting in irreversible drift. Non-rigid deformation of targets (e.g., pedestrian limb movement) challenges rigid-body assumptions, with part-level models still lacking dense correspondence to model fine-grained changes.

The accuracy-efficiency trade-off remains unresolved. Transformer-based architectures achieve superior performance but incur quadratic computational complexity with respect to point count, hindering real-time deployment on edge devices. Lightweight alternatives optimize speed but sacrifice precision in sparse scenes. Additionally, limited generalization across diverse scenarios (urban vs. off-road, varying sensor quality) persists, as models overfit to dataset-specific patterns rather than generalizable geometric or motion priors.