A Hybrid Model Based on CNN-MTF-LSTM Architecture Integrated Attention Mechanism for Vehicle Trajectory Prediction

MIMTP: Mamba-Driven Interaction-Aware Multi-Modal Trajectory Prediction for Autonomous Driving

An efficient Mamba-based feature extraction framework for jointly encoding vehicle trajectories and map information is proposed and achieves superior performance in terms of minADE, minFDE, and minMR, while maintaining high computational efficiency.

Vehicle Multimodal Trajectory Prediction Integrating Kinematics and Dynamic Interaction Features

Accurate vehicle trajectory prediction is essential for autonomous driving safety. However, existing data-driven models often ignore kinematic constraints, causing lateral jitter and trajectory distortion, while purely kinematics-based models lack flexibility in complex interactions. To address this, this paper presents a multimodal trajectory prediction method combining kinematics with dynamic interaction features. Operating in the Frenet coordinate system, the proposed model extracts historical features via a Bidirectional Gated Recurrent Unit (Bi-GRU) and utilizes an Adaptive Social Gating Network (ASGN) with multi-head attention to filter irrelevant interaction noise. This paper introduces a Stochastic Gating Decoder for multimodal latent variable sampling, adaptively fusing kinematics and data-driven paths to capture driver intention uncertainty while maintaining kinematic consistency. The model is trained using a composite loss function (Focal Loss and Best-of-K) to mitigate dataset long-tail distribution and trajectory divergence. Experiments on the HighD dataset show the proposed model achieves a minADE of 0.425 m and a minFDE of 0.955 m, outperforming baselines and reducing Lat-ADE by 53.9% compared to Social-GAN. These results confirm the model generates smoother, kinematically interpretable trajectories with higher accuracy in long-tail lane-changing scenarios.

Research on LSTM-based spatial target trajectory forecasting enhanced by attention mechanisms

To address the strong dependence of space object orbit prediction on physical models and initial conditions, as well as the difficulty of completely eliminating prediction errors, this study proposes a satellite orbit prediction correction method that integrates an attention mechanism with a long short-term memory (LSTM) network. Taking the LAGEOS satellite as the research object, the proposed method uses position error, velocity, and acceleration features extracted from historical orbital data to train a deep learning model for predicting one-day-ahead orbital errors and correcting the SGP4 orbit prediction results. The experimental results show that the ATLSTM model outperforms the LSTM, support vector machine (SVM), back propagation neural network (BP), and bidirectional long short-term memory (BiLSTM) models in both orbital error prediction and correction. The residual ratios of ATLSTM in the X, Y, and Z axes are reduced to 3.68%, 4.77%, and 2.37%, respectively, effectively improving the accuracy of satellite orbital error prediction. Further analysis indicates that a reasonable setting of the number of neurons helps improve model performance, while the prediction difficulty increases with the extension of the prediction duration, suggesting that the ATLSTM model is more suitable for short-term orbital error prediction and correction. In addition, validation results for satellites at different orbital altitudes demonstrate that the proposed model has certain generalization capability. In summary, combining deep learning methods with physical orbital models can effectively improve the accuracy of space object orbit prediction and provides an effective approach for orbital error prediction, space situational awareness, and collision warning.

A multitask CNN-LSTM model for joint prediction of vehicle lanechange intention and collision risk

Accurate prediction of vehicle lane-change (LC) behavior and potential collision risk in highway scenarios is important for advanced driver-assistance systems (ADAS). To address the separation between lane-change prediction and risk assessment in existing studies, the difficulty of identifying long-tailed high-risk scenarios, and the tendency of selfattention models to overlook local high-frequency hazard signals, this paper proposes a multi-task learning model based on convolutional neural networks and long short-term memory networks (CNN-LSTM). The model simultaneously predicts vehicle lane-change intention, time to lane crossing (TTLC), and collision risk level. Based on HighD naturalistic driving data, kinematic and interaction features of the target vehicle and surrounding vehicles are extracted; 1D-CNN is used to extract local spatial features, and a unidirectional LSTM is combined to capture temporal dependencies. Experimental results show that the proposed model achieves a lane-change intention prediction accuracy of 94.49%; in the high-risk classification task, its recall reaches 65.52%, outperforming Transformer (44.83%) and SVM (24.14%). The results indicate that the multi-task CNN-LSTM can balance macroscopic behavior prediction and microscopic risk recognition, thereby improving the active warning capability of autonomous-driving systems in complex traffic scenarios.

HCAD-Net: end-to-end parking network with historical context and attention-based dual-decoder

Autonomous parking requires accurate perception, reliable trajectory generation, and physically feasible vehicle motion in highly constrained environments. Existing end-to-end parking methods mainly rely on single-frame scene understanding and often neglect temporal motion priors and vehicle kinematic characteristics, limiting their planning accuracy and deployment robustness. To address these limitations, this paper proposes a vision-based end-to-end autonomous parking framework trained through imitation learning. The proposed framework introduces a historical context fusion encoder to capture temporal dependencies from past vehicle motions, a dual-stream attention decoder to enhance interaction between scene features and trajectory representations, and kinematic-aware auxiliary losses to enforce smooth and feasible trajectory generation. These components jointly improve trajectory prediction accuracy while maintaining vehicle manoeuvrability in parking scenarios. Extensive experiments demonstrate that the proposed method achieves over 25% lower trajectory prediction error than the baseline model on the ParkingE2E dataset and attains an 89.84% parking success rate in closed-loop CARLA simulations. The results verify the effectiveness of combining temporal context modeling, attention-based feature interaction, and kinematic constraint learning for end-to-end autonomous parking.

W2W: Language-Model-Based Trajectory Prediction with Reinforcement Learning