Deployment-friendly Lane-changing Intention Prediction Powered by Brain-inspired Spiking Neural Networks

TL;DR

The paper tackles real-time lane-changing intention prediction for autonomous driving by addressing deployment constraints of existing models. It introduces a brain-inspired Spiking Neural Network that encodes vehicle state time-series with a Linear feature expansion, a Leaky Integrate-and-Fire (LIF) temporal model, and a Softmax classifier to output three lane-change categories, operating on inputs X ∈ $\mathbb{R}^{[12,5]}$. Empirical results on the HighD and NGSIM datasets show substantial efficiency gains (e.g., training time reduced by up to 75% and memory by up to 99.9%) while achieving prediction accuracy comparable to LSTM and superior to ESN, with ROC-AUC values around 0.99+. The work demonstrates the practical potential of SNNs for deployment on resource-constrained autonomous driving systems, offering fast inference, low memory footprint, and energy-efficient operation, and sets the stage for broader real-world validation.

Abstract

Accurate and real-time prediction of surrounding vehicles' lane-changing intentions is a critical challenge in deploying safe and efficient autonomous driving systems in open-world scenarios. Existing high-performing methods remain hard to deploy due to their high computational cost, long training times, and excessive memory requirements. Here, we propose an efficient lane-changing intention prediction approach based on brain-inspired Spiking Neural Networks (SNN). By leveraging the event-driven nature of SNN, the proposed approach enables us to encode the vehicle's states in a more efficient manner. Comparison experiments conducted on HighD and NGSIM datasets demonstrate that our method significantly improves training efficiency and reduces deployment costs while maintaining comparable prediction accuracy. Particularly, compared to the baseline, our approach reduces training time by 75% and memory usage by 99.9%. These results validate the efficiency and reliability of our method in lane-changing predictions, highlighting its potential for safe and efficient autonomous driving systems while offering significant advantages in deployment, including reduced training time, lower memory usage, and faster inference.

Figures (5)

• Figure 1: Schematic of the SNN model for lane-changing intention prediction. The model processes a time-series input of vehicle state features through four main components: (1) Event Data Construction to generate the input matrix, (2) Feature Extraction with a Linear layer to expand feature dimensions, (3) Temporal Modeling using a LIF layer to capture temporal dependencies, and (4) Classification where a Linear layer maps the features to lane-change intention categories, followed by a Softmax activation for prediction.
• Figure 2: Comparison of average training time per epoch for our approach, LSTM, and ESN on the HighD and NGSIM datasets.
• Figure 3: Loss curves for SNN, ESN, and LSTM on the HighD (a) and NGSIM (b) datasets
• Figure 4: Comparison of lane-changing intention predictions for SNN, LSTM, and ESN models in the HighD (a) and NGSIM (b) datasets. The blue, red, and yellow curves represent predictions from the SNN, LSTM, and ESN models, respectively. Our method predicts lane-changing accurately, with no false predictions.
• Figure 5: ROC curves for SNN, LSTM, and ESN on the NGSIM (a) and HighD (b) datasets. The blue, red, and yellow curves represent the SNN, LSTM, and ESN methods, respectively.