Non-Autoregressive Sparse Transformer Networks for Pedestrian Trajectory Prediction

Abstract

Pedestrian trajectory prediction is an important task in practical applications such as automatic driving and surveillance systems. It is challenging to effectively model social interactions among pedestrians and capture temporal dependencies. Previous methods typically emphasized social interactions among pedestrians but ignored the temporal consistency of predictions and suffered from superfluous interactions by dense undirected graphs, resulting in a considerable deviance from reality. In addition, autoregressive approaches predicted future locations conditioning on previous predictions one by one, which would lead to error accumulation and time consuming. To address these issues, we present Non-autoregressive Sparse Transformer (NaST) networks for pedestrian trajectory prediction. Specifically, NaST models sparse spatial interactions and sparse temporal dependency via a sparse spatial transformer and a sparse temporal transformer separately. Different from previous predictions such as RNN-based approaches, the transformer decoder works in non-autoregressive pattern and predicts all the future locations at one time from a query sequence, which could avoid the error accumulation and be less computationally intensive. We evaluate our proposed method on the ETH and UCY datasets, and the experimental results show our method outperforms comparative state-of-the-art methods.

A Review of Trajectory Prediction Methods for the Vulnerable Road User

“…Apart from methods that use the original Transformer architecture, this chapter will also introduce methods that use an attention mechanism to predict trajectories. The Transformer architecture can be used without any changes to predict trajectories based on observed positions [17,58,59,[74][75][76].…”

“…Liu et al [59] used the embeddings of the last observed position as a query for the Transformer decoder part of the method, and Chen et al [74] supported the training of their Transformer structure with a multitask learning scheme that trained the model on intention recognition and trajectory prediction.…”

Synchronizing Object Detection: Applications, Advancements and Existing Challenges

“…A notable exemplar is EfficientDet, introduced in 2019, reflecting a trend towards optimizing the efficiency of object detection models [17]. Sparse attention mechanisms have emerged as a solution to handle largescale images effectively, enabling models to selectively attend to pertinent image regions, thereby reducing computation and improving overall efficiency [18]. The exploration of hybrid models, combining diverse architectures from twostage and one-stage detectors, along with the integration of ensemble methods, has shown promise in enhancing the overall performance of object detection systems [19].…”

Predicting Pedestrian Trajectories with Deep Adversarial Networks Considering Motion and Spatial Information

“…Many other works have also adopted this method [14,17,18,20,33], including the improved circular occupancy map [16] and angular grid [33]. Another non-grid method [24,26] embeds the relative position of other pedestrians with respect to the target pedestrian, similar to transformer position encoding [34,35], and then concatenates the relative position embeddings with the LSTM hidden states. The Multi-Layer Perceptron (MLP) and pooling layers then process the resulting vector to obtain the social vector.…”