Closed-Loop Integration of a Real-Time Modified OverFeat CNN Perception Engine with Model Predictive Control (MPC) for Precision Lane-Keeping and Collision Avoidance in Autonomous Guidance Systems

• Abbas Ahsun

  Texas University

  Author

The integration of robust perception systems with predictive control architectures remains a critical challenge in autonomous vehicle guidance, particularly for achieving reliable lane-keeping and collision avoidance under dynamic driving conditions. While convolutional neural networks (CNNs) have demonstrated exceptional performance in object detection and lane boundary identification, and model predictive control (MPC) has proven effective for trajectory optimization under constraints, their closed-loop integration introduces latency and stability challenges that compromise real-time performance. This study addresses this gap by proposing and validating a closed-loop architecture that integrates a modified OverFeat CNN perception engine operating at 10+ Hz with a nonlinear MPC framework for simultaneous lane-keeping and collision avoidance. The modified OverFeat CNN incorporates architectural modifications including skip-gram kernels for multi-scale context views, optimized bounding box merging algorithms to address object occlusion ambiguities, and lane boundary prediction capabilities. The perception engine provides real-time vehicle detection and lane geometry estimates that feed directly into the MPC's prediction horizon. Experimental validation demonstrates that the integrated system achieves 89.4% lane-keeping accuracy in dynamic scenarios and maintains a mean lateral deviation of 0.12 m across diverse driving conditions. The system successfully identifies and responds to potential collision threats with a detection-to-response latency of 120 ms, representing a 47% improvement over traditional cascaded architectures. This research contributes a validated framework for closed-loop perception-control integration and establishes benchmarks for real-time autonomous guidance system performance.

• pdf

Copyright (c) 2026 Abbas Ahsun (Author)

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