Spatial-Temporal Block and LSTM Network for Pedestrian Trajectories Prediction

1 Introduction

• Predicting pedestrian trajectories can support real-time decisions intended to prevent collisions with mobile platforms.The introduction also motivates early warning and rapid preventive measures.Source: §1 · p. 1
• Pedestrian decisions are uncertain, so trajectory prediction should be probabilistic rather than deterministic.Regression can average distinct intentions into a trajectory that does not fit the real scene.Source: §1 · p. 2
• The approach combines pedestrian positions, surrounding context, and previous trajectories as data-driven inputs.The model is designed to predict all pedestrian trajectories in crowd scenarios simultaneously.Source: §1 · p. 2
• The proposed ST-Block captures global interactions among pedestrians and local interactions with static objects, while graph embedding and LSTM model trajectories.The authors report better prediction than state-of-the-art methods.Source: §1 · p. 2

2 Related Works

• Prior approaches to human-space and human-human interactions may not model scene interactions or open-ended, unknown, or time-varying goals.The paper frames its methods for scenarios where people interact while dynamically anticipating future trajectories.Source: §2 · p. 2

2.1. LSTM networks for sequence prediction

• LSTM-based methods model sequence behavior, but some prior models omit scene context or predict a single future path rather than a distribution.The review also describes methods that share neighboring hidden states, model intention, or produce multiple destination-conditioned trajectories.Source: §2.1 · p. 3

2.2. Encoder-Decoder framework

• Encoder-decoder methods encode observed sequences and decode predictions, with prior variants using VAE or attention mechanisms.The review notes that naïve recurrent models may fail to model pedestrian interactions correctly.Source: §2.2 · p. 3

2.3. Spatial Temporal Graph Convolutional Networks

• The paper identifies few prior efforts combining spatial-temporal modeling and LSTMs for pedestrian sequences, and describes its ST-Blocks linking features with LSTM interaction modeling.The review contrasts this with earlier spatio-temporal graph work on traffic flow and human motion.Source: §2.3 · p. 4

2.4. Graph Embedding

• Graph embedding maps graph structure and node content into a compact feature space; this paper draws on a framework encoding both for scene features.The review notes that Social-STGCNN also extracts spatial and temporal information from graphs.Source: §2.4 · p. 4

3 Method

• The method combines graph and temporal convolutional features with interconnected LSTMs and a convolutional module to share information across pedestrians.The stated aim is to jointly reason about multiple people and predict their future trajectories simultaneously.Source: §3 · p. 4

3.1. Problem Definition

• The problem is to predict future trajectories for N pedestrians across T frames in a scene.The formulation distinguishes observed input trajectories from future ground-truth trajectories over specified time steps.Source: §3.1 · p. 4

3.2. Spatio-Temporal graph network

• The spatio-temporal network uses graph convolutions to extract spatial neighborhood features and gated CNNs to extract temporal dynamics.The graph convolution iteratively aggregates neighborhood embeddings; temporal convolutions form hierarchical representations and support parallel, controllable training.Source: §3.2 · pp. 4–5

3.3. Graph Representation of Pedestrian Trajectories

• At each time step, the graph represents pedestrian locations with nodes and connecting edges, alongside node attributes.The graph is written as G={Vt, Et, Xt}.Source: §3.3 · p. 5

• GCN extends convolution to graph data in the spectral domain, learning layer-wise transformations with a spectral convolution function.Source: §3.4 · p. 5
• Each convolution layer aggregates node and feature information, while its filter parameters are learned by the network.Source: §3.4 · p. 5
• A two-layer GCN uses an adjacency matrix augmented with self-connections and produces graph embedding H, with ReLU in the first layer and linear activation in the second.Source: §3.4 · p. 5

• TCN takes stacked sequential data as input and predicts a sequence as a whole.Source: §3.5 · p. 5
• Its 1D fully convolutional architecture preserves sequence length with zero padding, and causal convolutions at time t use only time t and earlier inputs.Source: §3.5 · p. 5
• The temporal model focuses on salient scene regions and more relevant neighboring pedestrians when predicting future states.Source: §3.5 · p. 5

3.6. Spatio-temporal Convolutional Block

• The model consistently outperformed baselines in FDE, with the authors attributing this to combining the ST-graph and LSTMs.Source: Table 1 · p. 6, §3.6 · p. 6
• ST-Blocks place a spatial layer between temporal layers to bridge graph convolution and temporal processing, compressing scale and squeezing features; layer normalization is used to prevent overfitting.The block can stack or extend graph features according to scale and complexity, and has three-dimensional input and output tensors containing batch size, maximum nodes, and sequence length.Source: §3.6 · p. 6

3.7. LSTM based Encoder-Decoder framework

• The encoder-decoder uses LSTM networks to process an input sequence into outputs iteratively, with the output computed from the hidden state.The output equation includes a projection matrix and bias vector.Source: §3.7 · p. 6

• LSTM networks encode each pedestrian's ST-block information and capture dependencies among ST-blocks from different states.Source: §3.8 · p. 6
• The encoder learns pedestrians' spatial and temporal states and stores their trajectory histories.Source: §3.8 · p. 6
• The LSTM encoder weights are shared across all people in a scene.Source: §3.8 · p. 6

3.9. Decoder module

• LSTM memory decodes pedestrians’ spatial and temporal features, reconstructs graph data, and predicts node locations.The paper describes bivariate Gaussian samples as a way to represent pedestrian positions.Source: §3.9 · p. 6
• The model predicts globally consistent, socially acceptable trajectories that stay closer to ground truth and avoid collisions by changing direction; the final column shows failures.The examples cover following, group avoidance, and merging scenarios.Source: §3.9 · p. 7, Figure 6 · p. 7
• In each BIWI/ETH scene, a collision is detected when the Euclidean distance between two pedestrians is less than 0.10 m; Table 2 reports average % of colliding pedestrians per frame.Source: §3.9 · p. 7, Table 2 · p. 7
• The prediction layer uses graph structure and content to determine node interactions, with pairwise edge probabilities contributing to graph reconstruction.The graph-construction error is minimized.Source: §3.9 · p. 7

4.1 Quantitative Evaluation

• Evaluation uses ETH and UCY trajectories, with 8 observed positions from 3.2 seconds to predict 12 future positions over 4.8 seconds.ADE averages L2 distance over predicted steps; FDE measures distance between predicted and true final destinations.Source: §4 · pp. 7–8, §4.1 · p. 8
• 0.40 ADE and 0.72 FDE are reported for our model, about 9.1% and 4% lower than Social-STGCNN’s 0.44 ADE and 0.75 FDE, respectively.The authors attribute the lower minimum FDE relative to S-GAN, SR-LSTM, and SoPhie to spatial and temporal feature awareness.Source: §4.1 · p. 9
• For near-collisions defined by pedestrian distances below 0.10 m, the model’s average collision percentage is reported as much better than three baselines.The comparison averages near-collisions across BIWI/ETH scenes using 30 random test agents and the first 8 seconds of each trajectory.Source: §4.1 · p. 9

4.2 Qualitative Analysis

• In following scenes, predictions are moderately accurate because the ground truth is closest to a linear path, and the authors describe their prediction as least different from actual movement.The passage notes that followers may continue following or speed up beyond the person ahead.Source: §4.2 · p. 9
• In group-avoidance scenes, S-LSTM and SGAN predictions show slight collisions, while ST-GCNN and the proposed model show desirable predictions.The passage links the less desirable predictions to insensitivity to neighboring pedestrians.Source: §4.2 · p. 9
• In merging scenes, the proposed model’s predictions match the real path, although some samples still show collision or divergence in the final column.S-LSTM and S-GAN trajectories show obvious collisions, while Social-GCNN deviates from the true path despite avoiding collision.Source: §4.2 · p. 9
• Across four collision-avoidance scenarios, 300 samples per setup are used to compare the proposed model with Social-GCNN through trajectory distributions and average predictions.The authors report that ST-Block adjusts the advancing direction to avoid possible collisions better than Social-GCNN in all four scenes.Source: §4.2 · p. 10

5 Conclusion

• Traditional methods are described as inadequate for mid- and long-term prediction because they often neglect spatial and temporal dependencies.Source: §5 · p. 10
• ST-Block captures adjacent graph relations while employing an LSTM network.Source: §5 · p. 10
• The authors report efficacy in several complicated real-life scenarios where social norms must be followed.Source: §5 · p. 10