Lightweight Spatiotemporal Highway Lane Detection via 3D-ResNet and PINet with ROI-Aware Attention
Authors: Sorna Shanmuga Raja, Abdelhafid Zenati
Organizations: Engineering Department, School of Science and Technology (SST), City, St George’s University of London, Northampton Square, London, EC1V 0HB, United Kingdom.
This paper presents a lightweight, end-to-end highway lane detection architecture that jointly captures spatial and temporal information for robust performance in real-world driving scenarios. Building on the strengths of 3D convolutional neural networks and instance segmentation, we propose two models that integrate a 3D-ResNet encoder with a Point Instance Network (PINet) decoder. The first model enhances multi-scale feature representation using a Feature Pyramid Network (FPN) and Self-Attention mechanism to refine spatial dependencies. The second model introduces a Region of Interest (ROI) detection head to selectively focus on lane-relevant regions, thereby improving precision and reducing computational complexity. Experiments conducted on the TuSimple dataset (highway driving scenarios) demonstrate that the proposed second model achieves 93.40% accuracy while significantly reducing false negatives. Compared to existing 2D and 3D baselines, our approach achieves improved performance with fewer parameters and reduced latency. The architecture has been validated through offline training and real-time inference in the Autonomous Systems Laboratory at City, St George's University of London. These results suggest that the proposed models are well-suited for integration into Advanced Driver Assistance Systems (ADAS), with potential scalability toward full Lane Assist Systems (LAS).
Figures & tables
Fig. 1 : Proposed Network 1.
Fig. 2 : Proposed Network 2.
Attribute
Details
Training Clips
3,626
Testing Clips
2,782
Frames per Clip
20
Resolution
1280×720 pixels
Annotations
Lane point coordinates (JSON format)
Scenario
Multi-lane highway driving scenes
TABLE I : Overview of TuSimple Dataset
Fig. 3 : (a) Lane annotation overlay, (b) mask visualisation.
Mechanism
Description
Limitation
Spatial Attention
Emphasises informative spatial regions
Limited long-range dependency modelling
Channel Attention
Reweights feature channels
Does not explicitly model spatial relationships
Self-Attention
Models global feature interactions
Higher computational complexity
TABLE II : Comparison of Attention Mechanisms
Layer
Resolution
Information Captured
1st (shallow)
128×256
Fine-grained spatial structures and edge-level details
2nd
64×128
Mid-level texture and structural representations
3rd
32×64
Higher-order feature abstractions and pattern compositions
4th (deepest)
16×32
High-level semantic representations and contextual encoding
TABLE III : Progression of Feature Maps in FPN
Fig. 4 : PINet implementation with multi-branch outputs.
Fig. 5 : ROI implementation and integration with loss functions.
Fig. 6 : Post-processing pipeline with RANSAC and graph-based smoothing.
Exp.
Model
Enhancements
1
3D-ResNet + PINet (Base Arch.)
Baseline configuration with RANSAC post-processing and graph-based smoothing
2
+ Self-Attention + FPN
Multi-scale feature fusion with multi-branch supervision
3
+ ROI + Focal + LineIoU
ROI-driven refinement with focal loss optimisation and improved lane alignment consistency
TABLE IV : Summary of Experimental Configurations
Fig. 7 : Results of Experiment 1: lane predictions, training stability, and confusion matrix.
Fig. 8 : Results of Experiment 2: lane predictions, training stability, and confusion matrix.
Fig. 9 : Results of Experiment 3: lane predictions, training stability, and confusion matrix.
Fig. 10 : Results of Experiment 4: lane predictions, training stability, and confusion matrix.