BEV-IMTP is presented, a lightweight instance mapping and trajectory prediction network designed for Bird's-Eye-View representations that achieves a 4.6% improvement in semantic map mIoU and a 1.8% gain in instance motion mIoU compared to state-of-the-art methods.
Abstract
Trajectory prediction is a critical component of autonomous driving systems, enabling vehicles to anticipate the motion of surrounding agents and make safe, informed navigation decisions. This paper presents BEV-IMTP, a lightweight instance mapping and trajectory prediction network designed for Bird’s-Eye-View representations. The proposed framework adopts a streamlined architecture with customized core layers and a minimal-parameter backbone, significantly reducing computational overhead while making it a promising candidate for latency-sensitive autonomous driving applications. BEV-IMTP achieves a 4.6% improvement in semantic map mIoU and a 1.8% gain in instance motion mIoU compared to state-of-the-art methods. On the nuScenes benchmark, the proposed model attains an overall semantic map mIoU of 62.1%, with strong class-wise performance of 63.4% for divider, 58.3% for pedestrian crossing, and 64.6% for boundary regions. Despite its high accuracy, BEV-IMTP remains computationally efficient, requiring only 34.5 million parameters and 88 GFLOPs, and supports inference at 3.1 FPS on a single Tesla V100 GPU. The model is trained on the nuScenes dataset and further evaluated on the Lyft dataset, demonstrating robust generalization across datasets captured using cameras with varying configurations. Extensive experiments on publicly available trajectory prediction benchmarks validated the effectiveness and efficiency of the proposed BEV-IMTP framework.
Trajectory prediction is essential for autonomous driving to anticipate traffic participants’ movements. While recent methods incorporate road topology, they often rely on heavy architectures, limiting real-time deployment. This study proposes a lightweight adversarial framework for multimodal trajectory prediction, integrating spatio-temporal encoding of agents and lanes, hierarchical attention-based scene fusion, a Bézier-parameterized generator, and a context-aware discriminator. This approach captures agent-road interactions and ensures trajectory realism with only 1.9M parameters. Experimental results demonstrate the superior speed and accuracy of the proposed method compared to existing approaches for trajectory prediction. The proposed method significantly enhances prediction accuracy, achieving results of minADE
6
0.76 and minFDE
6
1.12 on the Argoverse 1 dataset. In comparison to the state-of-the-art baseline model, there are notable improvements in minADE
6
and minFDE
6
by 0.01 and 0.03, respectively. Furthermore, our method generalizes well to the Argoverse 2 dataset, showing strong performance in more diverse and challenging environments.
Jiaxin Pang, Changshui Wu· Proceedings of the Instituti...· 0 citations
Existing visual navigation models often aim to develop foundation models that can generalize robot navigation across diverse platforms. However, many of these models are prone to collisions when deployed in complex indoor environments, particularly in structured layouts and narrow passages. To address this problem, we propose a depth image- and point-goal-conditioned navigation system, DevGRU. The proposed system employs an action predictor (AP) that generates collision-aware future trajectories, enabling effective avoidance of immediate obstacles. In conjunction with a collision predictor, the AP further compensates for errors accumulated in the goal pose estimation and proactively mitigates future deviations. To evaluate our method, we conducted experiments across nine different scenes and three state-of-the-art approaches - ViNT, NoMaD, and NavDP - as well as four additional variants of ViNT and NoMaD. In terms of navigation performance, DevGRU significantly outperforms ViNT and NoMaD by a large margin. In addition, the proposed model has a relatively small number of trainable parameters, resulting in the fastest inference time among the baselines, particularly outperforming NavDP by 7x in model size and 17x in inference time.
Kyung Min Han, Eunsom Kim, Young J. Kim· IEEE Robotics and Automation...· 0 citations
Understanding the real-time behaviour of pedestrians and their surrounding context is essential for road safety. Unfortunately, current perception systems struggle to capture how multiple cues jointly influence behavior in diverse traffic scenarios. Most existing models in Intelligent Transportation Systems (ITS) and Autonomous driving system (ADS) predict only a single attribute, such as intent or trajectory, while overlooking the interaction between behavioral and environmental factors. We propose a unified multi-attribute framework based on a Vision Transformer, which is enhanced with lightweight, parameter-efficient adapters. This innovative modular design enables independent learning of eight complementary attributes, including pedestrian behavior (action, attention, crossing, occlusion) and scene context (weather, time of day, road presence, density). Our proposed framework updates fewer than 2% of the model parameters. We ran our proposed framework using the JAAD dataset, and the framework achieves over 90% accuracy on most attributes and exhibits consistent behavior–context relationships and demonstrates robustness under varied environmental conditions. To evaluate real-time suitability, we implemented an adaptive strategy that activates only the most relevant adapters based on scene conditions. This approach reduces computational cost by 28.5% and increases throughput by 40.8% without loss of accuracy. This research distinctly highlights the practical potential of modular, PEFT-enhanced transformers for scalable multi-attribute perception as foundational vision models continue to advance. The best-performing adapters and inference code are available on Github Repo.
Accurate trajectory prediction is critical for autonomous driving safety and energy-efficient motion planning in sustainable urban mobility. This study isolates the effect of local lane graph conditioning by proposing a waterflow method that extracts an ego-centric lane topology via breadth-first traversal of the HD map (For convenience, all acronyms used throughout this manuscript (e.g., HD, LSTM, ADE, FDE, minADE, minFDE, BFS, GNN, SOTA) are collected in the Abbreviations section at the end of the paper), fusing lane features into trajectory encoders through cross-attention. The evaluation is deliberately scoped to ego-vehicle prediction at signal-controlled intersections: we evaluate across two architectures (LSTM and Transformer), two horizons (3 s and 8 s), and both single- and multi-modal (K=6) settings on 89,258 such scenarios from the Waymo Open Motion Dataset, so that “generality” refers to consistency across architectures, horizons, and output settings within this scope, rather than across prediction tasks. Lane conditioning consistently improves accuracy: +9.3% ADE at 3 s (p=0.007, 3 seeds), +26.6% minADE at 8 s (K=6, p=0.003, 3 seeds), and +26.8% ADE for the Transformer (p=0.030, 3 seeds)—with only ∼8% additional parameters for the LSTM. A controlled full-graph ablation (nearest 64 lanes) shows a consistent but not statistically significant trend favouring topologically guided local selection over brute-force spatial proximity (+11.4% minADE, p=0.063). Error decomposition reveals balanced lateral (+26.5%) and longitudinal (+25.4%) improvements, and a per-maneuver analysis over all 13,388 validation scenarios shows the largest gains for turning maneuvers. The lane-conditioned model (<700,000 parameters) runs in 0.8 ms per prediction on a desktop GPU (0.7 ms single-threaded CPU) with below 30 MB peak memory and an estimated 51 mJ per prediction, suggesting feasibility for resource-constrained deployment, pending validation on production automotive hardware.
Xing-Nan Zhou, C. Alecsandru· Sustainability· 0 citations
Bird’s-eye-view (BEV) perception has become the dominant paradigm for camera-centric scene understanding in autonomous driving, as well as road occupancy prediction, which involves the dense estimation of which regions of space are occupied and by what has emerged as its most expressive form. Between 2020 and 2026, the field underwent three overlapping transitions: from two-dimensional BEV semantic map segmentation to dense three-dimensional voxel-based 3D semantic occupancy, catalyzed by the 2022 industrial adoption of “occupancy networks,” and, most recently, to efficient, generative, and four-dimensional forecasting formulations. This survey organizes the literature along six orthogonal axes output representation, view-transformation mechanism, input modality, supervision paradigm, temporal scope, and efficiency strategy and uses the representation lineage as a primary spine connecting the 2020 BEV-segmentation works to the 2026 Gaussian and 4D frontier. Alongside the ego-centric mainstream, we review the parallel multi-view and infrastructure-side lineage from multi-view pedestrian occupancy to roadside traffic occupancy, which shares the BEV occupancy-map output and contributes generalization tools the ego-centric thread has yet to absorb. We review the canonical methods at each stage, summarize the standard datasets (CARLA, GMVD, MultiviewX, WildTrack, nuScenes, SemanticKITTI, Occ3D, OpenOccupancy) and evaluation metrics (MODA, mIoU, RayIoU, RayPQ), and consolidate reported results on the Occ3D-nuScenes benchmark into a single comparison. We close by identifying open problems in label efficiency, robustness, temporal forecasting, and deployment. Our intent is to bridge the historically separate BEV-segmentation and 3D-occupancy literatures within a single taxonomy.
Abdelrahman S. Heikal, Mostafa Farouk Senussi, Ahmed Salem et al.· Mathematics· 0 citations
Accurate pedestrian trajectory prediction is fundamental for safe navigation of autonomous robots and vehicles in crowded environments. Although graph-based methods such as Social-STGCNN efficiently model social interactions among pedestrians, they largely ignore static environmental constraints such as walls and obstacles, which can lead to physically infeasible predictions. In this paper, we propose an extended trajectory prediction method that integrates local and global environmental information into Social-STGCNN via Multi-Head Cross-Attention, and incorporates environmental constraint learning through contrastive MapNCE and collision avoidance EnvCol losses. To generate diverse prediction while maintaining scene consistency, we further introduce a low-dimensional trajectory representation based on Singular Value Decomposition and Adaptive Anchors derived from K-means clustering. We evaluate the proposed method on the ETH/UCY benchmark across five scenes using Average Displacement Error (ADE), Final Displacement Error (FDE), obstacle collision rate, and inference time. The results show that the proposed method consistently improves the FDE over Social-STGCNN, while ADE increases. The obstacle collision rate also increases, revealing a trade-off between endpoint accuracy and full-trajectory environmental compliance.
Yuka Takahara, Yuka Kato· International Conference on...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
