DenseTNT: End-to-end Trajectory Prediction from Dense Goal Sets

Gu, Junru; Sun, Chen; Zhao, Hang

doi:10.48550/arxiv.2108.09640

Cited by 3 publications

(7 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2) Metrics: We follow the official evaluation metrics of the Argoverse benchmark and focus on several important ones that are most considered in recent works [10,13,14,23]. The average displacement error (ADE) is the average Euclidean distance between the predicted and GT trajectories, while the final displacement error (FDE) only calculates the error at endpoints.…”

Section: Resultsmentioning

confidence: 99%

“…Then, final endpoints are sampled over the heatmap. The approach most related to ours is DenseTNT [23], which uses VectorNet to encode sparse context, and utilize an attention module to pass features from sparse nodes to all goal candidates, where the candidates are densely sampled w.r.t. a set of promising lanes.…”

Section: Related Workmentioning

confidence: 99%

“…After the dual-scale feature fusion, we perform the pernode classification on the DA graph, obtaining a heatmap that describes the probability distribution of the potential goal. Given the heatmap, some methods utilize a postprocessing module to generate final predicted goals, such as non-maximum suppression (NMS) [13] and optimizationbased algorithms [13,23]. Such methods usually depend on human heuristics and bring more effort in parameter tuning.…”

Section: G Multi-modal Goal Decodermentioning

confidence: 99%

“…Then, we investigate whether the fine-grained features are helpful for motion prediction by removing the feature aggregation process from the DA layer to the LS layer (DA2LS). Intuitively, the resulting variant is similar to DenseTNT [23], as both of the models fuse high-level features on a sparse graph and propagate features to dense endpoint samples to generate the goal heatmap. However, the differences are two-folds: 1) we use LaneConv for sparse feature aggregation while DenseTNT fuses features on a complete graph; 2) DenseTNT uses a global attention mechanism to pass messages from the sparse graph to dense nodes, while we use a distance-guided GAT, which is more efficient for computation and memory usage.…”

Section: Ablation Studymentioning

confidence: 99%

“…We can find that the NMS-based goal selection yields better coverage of the high-probability region due to its greedy property resulting in a lower MR, while our NN-based goal decoder directly optimizes minFDE, leading to lower displacement error. Previous works such as [13] and [23] propose postprocessing methods to further decrease the displacement error. However, these methods are always separated from the prediction network and are not differentiable.…”

Section: Ablation Studymentioning

confidence: 99%

See 4 more Smart Citations

Trajectory Prediction with Graph-based Dual-scale Context Fusion

Zhang

Chen³

et al. 2021

Preprint

View full text Add to dashboard Cite

Motion prediction for traffic participants is essential for a safe and robust automated driving system, especially in cluttered urban environments. However, it is highly challenging due to the complex road topology as well as the uncertain intentions of the other agents. In this paper, we present a graph-based trajectory prediction network named the Dual Scale Predictor (DSP), which encodes both the static and dynamical driving context in a hierarchical manner. Different from methods based on a rasterized map or sparse lane graph, we consider the driving context as a graph with two layers, focusing on both geometrical and topological features. Graph neural networks (GNNs) are applied to extract features with different levels of granularity, and features are subsequently aggregated with attention-based inter-layer networks, realizing better local-global feature fusion. Following the recent goaldriven trajectory prediction pipeline, goal candidates with high likelihood for the target agent are extracted, and predicted trajectories are generated conditioned on these goals. Thanks to the proposed dual-scale context fusion network, our DSP is able to generate accurate and human-like multi-modal trajectories. We evaluate the proposed method on the large-scale Argoverse motion forecasting benchmark, and it achieves promising results, outperforming the recent state-of-the-art methods.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: G Multi-modal Goal Decodermentioning

confidence: 99%

Section: Ablation Studymentioning

confidence: 99%

Section: Ablation Studymentioning

confidence: 99%

See 3 more Smart Citations

Trajectory Prediction with Graph-based Dual-scale Context Fusion

Zhang

Chen³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

Towards Traffic Scene Description: The Semantic Scene Graph

Zipfl

Zöllner

2022

2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC)

View full text Add to dashboard Cite

Autonomous vehicles hold great promise in improving the future of transportation. The driving models used in these vehicles are based on neural networks, which can be difficult to validate. However, ensuring the safety of these models is crucial. Traditional field tests can be costly, time-consuming, and dangerous. To address these issues, scenario-based closed-loop simulations can simulate many hours of vehicle operation in a shorter amount of time and allow for specific investigation of important situations. Nonetheless, the detection of relevant traffic scenarios that also offer substantial testing benefits remains a significant challenge. To address this need, in this paper we build an imitation learning based trajectory prediction for traffic participants. We combine an image-based (CNN) approach to represent spatial environmental factors and a graph-based (GNN) approach to specifically represent relations between traffic participants. In our understanding, traffic scenes that are highly interactive due to the network's significant utilization of the social component are more pertinent for a validation process. Therefore, we propose to use the activity of such sub networks as a measure of interactivity of a traffic scene. We evaluate our model using a motion dataset and discuss the value of the relationship information with respect to different traffic situations.

show abstract

Real-Time People Re-Identification and Tracking for Autonomous Platforms Using a Trajectory Prediction-Based Approach

Ghita

Florea

2022

Sensors

View full text Add to dashboard Cite

Currently, the importance of autonomous operating devices is rising with the increasing number of applications that run on robotic platforms or self-driving cars. The context of social robotics assumes that robotic platforms operate autonomously in environments where people perform their daily activities. The ability to re-identify the same people through a sequence of images is a critical component for meaningful human-robot interactions. Considering the quick reactions required by a self-driving car for safety considerations, accurate real-time tracking and people trajectory prediction are mandatory. In this paper, we introduce a real-time people re-identification system based on a trajectory prediction method. We tackled the problem of trajectory prediction by introducing a system that combines semantic information from the environment with social influence from the other participants in the scene in order to predict the motion of each individual. We evaluated the system considering two possible case studies, social robotics and autonomous driving. In the context of social robotics, we integrated the proposed re-identification system as a module into the AMIRO framework that is designed for social robotic applications and assistive care scenarios. We performed multiple experiments in order to evaluate the performance of our proposed method, considering both the trajectory prediction component and the person re-identification system. We assessed the behaviour of our method on existing datasets and on real-time acquired data to obtain a quantitative evaluation of the system and a qualitative analysis. We report an improvement of over 5% for the MOTA metric when comparing our re-identification system with the existing module, on both evaluation scenarios, social robotics and autonomous driving.

show abstract

DenseTNT: End-to-end Trajectory Prediction from Dense Goal Sets

Cited by 3 publications

References 32 publications

Trajectory Prediction with Graph-based Dual-scale Context Fusion

Trajectory Prediction with Graph-based Dual-scale Context Fusion

Towards Traffic Scene Description: The Semantic Scene Graph

Real-Time People Re-Identification and Tracking for Autonomous Platforms Using a Trajectory Prediction-Based Approach

Contact Info

Product

Resources

About