CRAT-Pred: Vehicle Trajectory Prediction with Crystal Graph Convolutional Neural Networks and Multi-Head Self-Attention

Schmidt, Jochen; Jordan, Julian; Gritschneder, Franz; Dietmayer, Klaus

doi:10.48550/arxiv.2202.04488

Cited by 1 publication

(8 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our social baseline is inspired in the architecture proposed by [36]. It uses as input the past trajectories of the most relevant obstacles as relative displacements to feed the Encoding Module (Fig.…”

Section: A Social Baselinementioning

confidence: 99%

“…2). Then, the social information is computed using a Graph Neural Network (GNN), in particular Crystal-Graph Convolutional Network (Crystal-GCN) layers [36], [41], and Multi-Head Self Attention (MHSA) [42] to obtain the most relevant agent-agent interactions. Finally, we decode this latent information using an autoregressive strategy where the output at the i-th step depends on the previous one for each mode respectively in the Decoding Module.…”

Section: A Social Baselinementioning

confidence: 99%

“…3) Social Attention module: After encoding the past history of each vehicle in the sequence, we compute the agent-agent interactions to obtain the most relevant social information of the scene. For this purpose, we construct an interaction graph using Crystal-GCN [41] [36]. Then, MHSA [42] is applied to enhance the learning of agent-agent interactions.…”

Section: A Social Baselinementioning

confidence: 99%

“…• Social information: In complex scenarios, motion history encoding of a particular target agent is not sufficient to represent the latent space of the traffic situation, but the algorithm must deal with a dynamic set of neighbouring agents around the target agent. Common techniques are aggregating neighbour motion history with a permutationinvariant set operator: soft attention [20], [33], a combination of soft attention and RNN [24] / GNN [36] or social pooling [19], [29]. Raster based approaches rely on 2D convolutions [14] [34] over the spatial grid to implicitly capture agent interactions in such a way longterm interactions are dependent on the neural network receptive fields.…”

Section: Introductionmentioning

confidence: 99%

“…• Decoder: Pioneering works of DL based MP usually adopt the autoencoder architecture, where the decoder is often represented by a recurrent network (GRU, LSTM, etc., specially designed to handle temporal information) to generate future trajectories in an autoregressive way, or by CNNs [8] [15] / MLP [16] [36] using the nonautoregressive strategy. The method may use an autoregressive strategy where the pipeline generates tokens (in this case, positions or relative displacements) in a sequential manner, in such a way the new output is dependent on the previously generated output, whilst MLP [36], CNN [8] or transformer [33] based strategies usually follow a non-autoregressive strategy, where from a latent space the whole future trajectory is predicted. • Output: The most popular model output representation is a sequence of states (absolute positions) or state differences (relative displacements for any dimension considered).…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Exploring Attention GAN for Vehicle Motion Prediction

Gómez-Huélamo¹,

Conde²,

Ortiz³

et al. 2022

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

View full text Add to dashboard Cite

Motion Prediction (MP) of multiple surroundings agents is a crucial task in arbitrarily complex environments, from simple robots to Autonomous Driving Stacks (ADS). Current techniques tackle this problem using end-to-end pipelines, where the input data is usually a rendered top-view of the physical information and the past trajectories of the most relevant agents; leveraging this information is a must to obtain optimal performance. In that sense, a reliable ADS must produce reasonable predictions on time. However, despite many approaches use simple ConvNets and LSTMs to obtain the social latent features, State-Of-The-Art (SOTA) models might be too complex for realtime applications when using both sources of information (map and past trajectories) as well as little interpretable, specially considering the physical information. Moreover, the performance of such models highly depends on the number of available inputs for each particular traffic scenario, which are expensive to obtain, particularly, annotated High-Definition (HD) maps.In this work, we propose several efficient baselines for the well-known Argoverse 1 Motion Forecasting Benchmark. We aim to develop compact models using SOTA techniques for MP, including attention mechanisms and GNNs. Our lightweight models use standard social information and interpretable map information such as points from the driveable area and plausible centerlines by means of a novel preprocessing step based on kinematic constraints, in opposition to black-box CNN-based or toocomplex graphs methods for map encoding, to generate plausible multimodal trajectories achieving up-to-pair accuracy with less operations and parameters than other SOTA methods. Our code is publicly available at https://github.com/Cram3r95/mapfe4mp.

show abstract

Section: A Social Baselinementioning

confidence: 99%

Section: A Social Baselinementioning

confidence: 99%