A Multi-Modal Deep Learning Architecture Integrating Modified OverFeat CNN with Solid-State LiDAR Point Clouds for Occlusion-Resistant Real-Time Vehicle and Lane Boundary Detection

• Abbas Ahsun

  Texas University

  Author

Reliable real-time perception remains a critical bottleneck for autonomous driving systems operating in dynamic, occlusion-prone environments. While vision-based detection methods have demonstrated significant progress, they exhibit fundamental vulnerabilities under adverse conditions characterized by partial occlusions, challenging lighting scenarios, and complex traffic interactions. This study proposes a novel multi-modal deep learning architecture that integrates a Modified OverFeat Convolutional Neural Network with Solid-State LiDAR point cloud data to achieve robust, occlusion-resistant vehicle and lane boundary detection. The proposed framework leverages the complementary strengths of visual semantic features from the Modified OverFeat architecture and geometric-spatial information from LiDAR point clouds through a mid-fusion strategy, incorporating attention mechanisms to dynamically weight modality contributions based on environmental context. The system was trained and evaluated on a comprehensive dataset comprising annotated frames from multiple sensors. The proposed architecture achieved a detection accuracy of 89.4% for vehicle detection and a lane boundary intersection over union (IoU) of 87.2%, demonstrating statistically significant improvements over single-modality baselines and state-of-the-art fusion approaches. The framework maintained real-time inference at 12 frames per second on edge GPU configurations, substantially reducing occlusion-induced detection failures. These findings establish a replicable architectural template for robust multi-modal perception in autonomous driving applications, with direct implications for safety-critical system design and deployment

• pdf

Copyright (c) 2026 Abbas Ahsun (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.