[2024107_LabSeminar_Huy]MFTraj: Map-Free, Behavior-Driven Trajectory Prediction for Autonomous Driving.pptx

MFTraj: Map-Free, Behavior-Driven Trajectory Prediction for Autonomous Driving

1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea 2024-10-07 MFTraj : Map-Free, Behavior-Driven Trajectory Prediction for Autonomous Driving Haicheng Liao et al. IJCAI '24: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence

2. OUTLINE MOTIVATION METHODOLOGY EXPERIMENT & RESULT CONCLUSION

3. MOTIVATION The integration of autonomous vehicles (AVs) with human driven vehicles and pedestrians necessitates advanced trajectory prediction models: predict the future trajectories of various road users, leveraging historical data . Overview and Limitation Challenges: modeling the often-unpredictable driving behaviors of road users The need for understanding the cognitive processes that dictate those path. AVs can anticipate sudden changes in human-driven vehicles or pedestrian movements, leading to safer co-navigation. Behavior-focused predictions can aid in scenarios where traditional data might be ambiguous or incomplete (missing values). High Definition (HD) maps is resource-intensive, which can become obsolete in rapidly changing environments even with map-free HD model. lack the granularity provided by comprehensive road network data.

4. INTRODUCTION A novel dynamic geometric graph . Captures essence of continuous driving behavior, circumventing limitations of manual labeling. Integrated behavioral metrics and criteria (traffic psychology, cognitive neuroscience, and decision-making framework). craft a model that offers more than mere predictions—it elucidates. Contribution An advanced map-free architecture for trajectory prediction : obviates the need for HD maps, resulting insignificant computational savings.

5. METHODOLOGY Inputs and Outputs At time t, the ego vehicle anticipates the target vehicle’s trajectory for the upcoming steps Past trajectories of both the target vehicle (indexed by 0) and its surrounding agents (indexed from 1 to n) over a predefined horizon Future trajectory of the target vehicle during the prediction horizon : is the 2D coordinates of the target vehicle at time t.  

6. METHODOLOGY Main Architecture Fig. Architecture of the proposed trajectory prediction model.

7. METHODOLOGY Behavior-aware Module Draws inspiration from the multipolicy decision-making framework integrating elements of traffic psychology and dynamic geometric graphs (DGGs) to effectively capture intricate driving behaviors amid ongoing driving maneuvers and evolving traffic conditions. DGGs: model the interactions of different agents(nodes) : graph at time t. An edge , where Adjacency matrix is defined from Euclidean distance.  

8. METHODOLOGY Behavior-aware Module Centrality Measures: provide valuable insights into the importance, influence, and connectivity of nodes or vertices within a graph. Degree Centrality: Closeness Centrality: Eigenvector Centrality: Betweenness Centrality: Power Centrality: Katz Centrality:

9. METHODOLOGY Behavior-aware Module Behavior-aware Criteria: evaluate different driving behaviors for target vehicle and its surroundings . Behavior Magnitude Index (BMI) , Behavior Tendency Index (BTI) , and Behavior Curvature Index (BCI) .   GRU and variational RNN functions Behavior Encoder: model relationships between random variables across time . captures human driving patterns and their temporal dynamics.

10. METHODOLOGY Main Architecture Fig. Overview of our adaptive structure-aware GCN.

11. METHODOLOGY Position-aware Module Emphasizes relative positions: captures individual and group spatial dynamics, interpreting the scene’s geometric nuances. positional features output by the position encoder Pooling Mechanism : captures dynamic position data from the traffic environment around the target vehicle . Position Encoder : transform discrete position vectors into continuous spatiotemporal representation enhancing temporal and spatial interactions between agents and map elements hidden position state updated on a frame-by-frame

12. METHODOLOGY Interaction-aware Module and Decoder Interaction-aware : capability to craft spatial feature matrices dynamically using GCN . Adjusting to the number of agents observed in real-time. Multifaceted spatio -temporal interactions + physical interplays between agents. Initial feature vector: Layer k Soft plus Sigmoid Element-wise product Residual Decoder: a linear residual and projection layer for future forecasting . Group Normalization (GN) function

13. EXPERIMENT AND RESULT Experiment Settings Dataset: Argoverse , NGSIM, HighD , and MoCAD . Baselines: Argoverse , Constant Velocity [ 1 ], SGAN[2], TPNet [3], PRIME[4], Uulm-mrm [1], WIMP[5], Scence -Transformer[6], CtsCov [7], HOME[8], LaneGCN [9], GOHOME[10], DenseTNT [11], VectorNet [12], TPCN[13], SSL-Lanes[14], LTP [15], HiVT-128 [16]. [1] Chang, Ming-Fang, et al. " Argoverse : 3d tracking and forecasting with rich maps." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.. [2] Gupta, Agrim , et al. "Social gan : Socially acceptable trajectories with generative adversarial networks." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. [3] Fang, Liangji , et al. " Tpnet : Trajectory proposal network for motion prediction." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. [4] Song, Haoran , et al. "Learning to predict vehicle trajectories with model-based planning." Conference on Robot Learning. PMLR, 2022. [5] Khandelwal, Siddhesh, et al. "What-if motion prediction for autonomous driving." arXiv preprint arXiv:2008.10587 (2020). [6] Ngiam , Jiquan , et al. "Scene transformer: A unified architecture for predicting multiple agent trajectories." arXiv preprint arXiv:2106.08417 (2021). [7] Zhao, Hang, et al. " Tnt : Target-driven trajectory prediction." Conference on Robot Learning. PMLR, 2021. [8] Gilles, Thomas, et al. "Home: Heatmap output for future motion estimation." 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. [9] Liang, Ming, et al. "Learning lane graph representations for motion forecasting." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. Springer International Publishing, 2020. [10] Gilles, Thomas, et al. " Gohome : Graph-oriented heatmap output for future motion estimation." 2022 international conference on robotics and automation (ICRA). IEEE, 2022. [11] Gu, Junru , Chen Sun, and Hang Zhao. " Densetnt : End-to-end trajectory prediction from dense goal sets." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. [12] Gao, Jiyang , et al. " Vectornet : Encoding hd maps and agent dynamics from vectorized representation." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020. [13] Ye, Maosheng , Tongyi Cao, and Qifeng Chen. " Tpcn : Temporal point cloud networks for motion forecasting." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [14] Bhattacharyya, Prarthana , Chengjie Huang, and Krzysztof Czarnecki. " Ssl -lanes: Self-supervised learning for motion forecasting in autonomous driving." Conference on Robot Learning. PMLR, 2023. [15] Wang, Jingke , et al. " Ltp : Lane-based trajectory prediction for autonomous driving." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. [16] Zhou, Zikang , et al. " Hivt : Hierarchical vector transformer for multi-agent motion prediction." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. Measurement : minADE , minFDE , MR, and RMSE.

14. EXPERIMENT AND RESULT Result – Overall Perfor mance Tab . Performance comparison on complete and missing datasets Tab . Comparative evaluation of MFTraj with SOTA baselines.

15. EXPERIMENT AND RESULT R esult – Visualization . Fig. Qualitative results of MFTraj on NGSIM. Fig. Ablation analysis of individual components in Argoverse .

16. CONCLUSION Presents a map-free and behavior-aware trajectory prediction model for AVs : 4 components: behavior-aware, position-aware, interaction-aware modules, and a residual decoder. work in concert to analyze and interpret various inputs, understand human machine interactions, and account for the inherent uncertainty and variability in the prediction. Through experiment result. underscores the resilience and efficiency of MFTraj in predicting future vehicle trajectories suggests its potential to drastically reduce the data requirements for training AVs, especially in data-missing and limited data scenes. Summarization