TridentNetV2: Real-time Graph Navigation

• TridentNetV2 is a dynamic trajectory generation framework that replaces HD maps with a lightweight, graph-based global plan derived from OSM and probabilistic semantic scene encoding.
• It employs a self-attention graph encoder and CVAE to reduce parameter count by 31% and deliver multi-modal trajectory predictions in real time, particularly at urban intersections.
• The framework fuses global planning and local semantic cues to enhance trajectory accuracy, compliance, and computational efficiency compared to traditional raster-based methods.

TridentNetV2 is a dynamic trajectory generation framework for autonomous navigation that eliminates reliance on centimeter-level-labeled high-definition (HD) maps. Instead, it leverages a graph-based global plan constructed automatically from OpenStreetMap (OSM) data and incorporates a lightweight, probabilistic semantic scene encoding. TridentNetV2 achieves real-time, multi-modal trajectory prediction with measurable gains in accuracy, compliance, and computational efficiency, particularly in urban intersection scenarios, while reducing network complexity compared with raster-based or HD-map–dependent approaches (Paz et al., 2022).

1. Core Contributions and Innovations

TridentNetV2 addresses the limitations of prior autonomous navigation pipelines dependent on intensive HD map annotation. The framework introduces several distinct advances:

• Graph-Structured Global Plan: Replaces rasterized or HD maps with a nominal, ego-centric global plan realized as a lightweight graph structure derived from OSM, encoding explicit road connectivity and intersection attributes (stop signs, traffic lights, crosswalks).
• Contextual Semantic Fusion: Merges the graphical plan with a probabilistically built local semantic map (driveable areas, sidewalks, lane markings, vegetation) into a hybrid representation, integrating contextual scene information.
• Self-Attention Graph Encoder: Encodes the global-plan graph via a Self-Attention architecture and MLP rather than a convolutional scheme over raster images, resulting in a 31% parameter reduction (11.5M vs 16.8M parameters) and removal of IMU-based drift evident in rasterized plan encoding.
• Conditional Variational Autoencoder (CVAE): Employs a CVAE whose discrete latent modes capture the maneuver multi-modality (e.g., left turn, right turn). This is augmented with a pointwise mean squared error (MSE) loss term to improve trajectory fidelity.
• Dataset Contribution: Releases IntersectionScenes1.0, a dataset dedicated to three- and four-way urban intersection maneuvers.

These innovations result in a framework capable of competitive accuracy relative to map-based solutions, at lower computational and data annotation cost.

2. Network Architecture and Scene Representation

The TridentNetV2 pipeline operates at 10 Hz, leveraging both a global planning and local semantic perception channel:

• Graph-Based Global Plan: Constructed from OSM, each instance comprises a sequence of $P=20$ past and $F=20$ future waypoints, for a total of 40 nodes. Every node $v_i$ is a tuple $(p_x, p_y, f)$, with $f$ denoting the node feature (planned traversed, stop sign, traffic light, or crosswalk).
• Scene Encoding: The local scene is encoded into a semantic map following [Paz et al. 2020], fused from camera and LiDAR, represented as a $400 \times 400 \times 3$ tensor $m_s$ that captures driveable areas, sidewalks, lane markings, vegetation, and crosswalks. The crop window covers 100 m at 2 px/m resolution.
• Model Components:
  • Self-Attention module and MLP encode $m_g$ into $h_g$
  • A compact 2D CNN encodes $m_s$ into $F=20$0
  • The concatenated embedding $F=20$1 is input to the CVAE prior $F=20$2
  • During training, ground truth trajectory $F=20$3 is additionally encoded via $F=20$4
  • A GRU-based decoder outputs, at each waypoint step $F=20$5, a bivariate Gaussian $F=20$6 for trajectory generation

Relevant architecture formulas include the self-attention operation on the graph nodes:

$F=20$7

where $F=20$8, $F=20$9, $v_i$0 are the learned projections and $v_i$1.

3. Mathematical Framework and Generative Modeling

TridentNetV2 models the conditional trajectory distribution $v_i$2, where $v_i$3 denotes the hybrid scene encoding:

$v_i$4

Here, $v_i$5 is a discrete latent maneuver mode, with $v_i$6. The CVAE structure is trained using a variational lower bound (ELBO) with a recognition model $v_i$7:

$v_i$8

An additional MSE loss term sharpens predicted trajectories:

$v_i$9

with $(p_x, p_y, f)$0 waypoints.

Inference is performed by selecting $(p_x, p_y, f)$1 and decoding via the GRU for $(p_x, p_y, f)$2 steps, yielding average network runtime of approximately 6 ms per update on GPU, corresponding to 6% of a 100 ms control cycle.

4. Datasets, Metrics, and Empirical Results

TridentNetV2 was evaluated on:

• NominalScenes1.0: Urban driving (UCSD campus), including lane following, intersection maneuvers, U-turns, and sharp turns. 6,128 training and 2,864 test sequences.
• IntersectionScenes1.0: Urban intersections (three- and four-way entries/exits), comprising 2,924 training and 1,506 test samples.

Evaluation Metrics:

Metric	Definition
ADE_HALF	Avg. displacement error over first $(p_x, p_y, f)$3 waypoints
ADE_FULL	Avg. displacement error over all $(p_x, p_y, f)$4 waypoints
FDE	Final displacement error at $(p_x, p_y, f)$5
MDE	Worst-case error across the predicted sequence
DAC_HALF	Fraction of first $(p_x, p_y, f)$6 waypoints in driveable area
DAC_FULL	Fraction of all $(p_x, p_y, f)$7 waypoints in driveable area

Quantitative Results:

Dataset	Baseline	ADE_FULL ↓	FDE ↓	MDE ↓	DAC_FULL ↑
NominalScenes1.0	Raster	1.056 m	2.448	2.495	0.849
	TridentNetV2	0.969 m	2.317	2.393	0.915
IntersectionScenes1.0	Raster	1.793 m	3.672	—	0.858
	TridentNetV2	1.511 m	3.088	—	0.898

Ablation Studies indicate that inclusion of stop sign and traffic light features (Graph-STPF) achieves optimal balance of accuracy and drivable-area compliance, while adding crosswalks (Graph-STCPF) can slightly degrade performance due to their occurrence on straight paths.

5. Computational Efficiency and Complexity Analysis

TridentNetV2 achieves a notable reduction in model complexity:

• Self-Attention Graph Encoder: 11.5M parameters (vs 16.8M for raster-based CNN)
• Total Runtime: ~6 ms per inference, supporting real-time embedded operation (occupying 6% of typical control cycle)

Parameter reduction is achieved primarily through replacement of convolutional processing on rasterized plans with attention-driven graph encoding. IMU-drift inherent to raster-based plan registration is averted by transforming graph coordinates into the ego-vehicle frame using OSM global heading.

6. Implications, Limitations, and Future Directions

Abandonment of HD maps in favor of OSM-driven graph planning and semantic map fusion enhances scalability, reduces dependency on frequent manual map annotation, and enables adaptation to unstructured road changes such as construction. The hybrid representation and attention mechanism facilitate robust, multi-modal trajectory generation without the incremental error associated with inertial navigation.

Identified limitations of TridentNetV2 include:

• Absence of explicit kinematic or vehicle-dynamic constraint modeling within the generative trajectory model.
• Inability in its current form to jointly predict future behaviors of other dynamic agents (e.g., pedestrians, cyclists) as explicit nodes in the planning graph.
• Reliance on preconstructed local semantic maps, presenting challenges in entirely novel scenes; incorporation of online mapping updates remains an open area.

This suggests that further work could focus on integration of dynamic agent modeling and real-time map adaptation mechanisms.

7. Summary and Context in Autonomous Navigation Research

TridentNetV2 represents a substantive step towards scalable, real-time generative planning in urban autonomous navigation without high-definition map overhead. By combining a graph-encoded OSM global plan, contextual semantic scene mapping, and a CVAE-driven prediction architecture, the framework delivers multi-modal, compliant, and accurate trajectories benchmarked across intersection-rich urban datasets. The approach provides a template for subsequent research oriented around lightweight, context-aware planning systems and robust generative trajectory modeling, competitive with HD-map–centered methods but with significantly relaxed infrastructure prerequisites (Paz et al., 2022).