Optimization of Lightweight Modified OverFeat Convolutional Neural Networks for Edge-Computing Real-Time Multi-Task Perception in Resource-Constrained Autonomous Vehicles

• Abbas Ahsun

  Texas University

  Author

The deployment of real-time perception systems for autonomous vehicles faces fundamental challenges in balancing computational efficiency, detection accuracy, and multi-task capability within stringent edge-computing constraints. While convolutional neural networks have demonstrated remarkable performance in vehicle and lane detection, existing architectures such as the OverFeat CNN require substantial computational resources that limit their viability on resource-constrained embedded platforms. This study addresses this gap by proposing a lightweight modified OverFeat CNN framework optimized for edge-computing real-time multi-task perception. The methodology integrates channel sparsification techniques to reduce model parameters by 62.3%, network pruning to compress the architecture while maintaining feature representation quality, and quantization-aware training to enable efficient inference on edge devices. The proposed framework is evaluated on the BDD100K dataset under diverse driving conditions, achieving vehicle detection accuracy of 89.4% mAP, lane detection accuracy of 87.2%, and inference speeds of 23.7 FPS on NVIDIA Jetson Xavier NX while consuming 7.8W. Comparative analysis demonstrates a 2.3× speedup and 41.5% reduction in energy consumption against baseline OverFeat implementations, with performance comparable to dual-network architectures at significantly reduced computational cost. These findings provide a replicable framework for edge-deployable perception systems, contributing practical guidance for autonomous vehicle system designers seeking real-time multi-task perception under power and thermal constraints.

• pdf

Copyright (c) 2026 Abbas Ahsun (Author)

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