BALANCING PERFORMANCE AND CONSTRAINTS IN AUTONOMOUS DRIVING: A SYSTEMATIC REVIEW OF CNN TECHNIQUES, SAFETY VALIDATION, AND OPERATIONAL ROADBLOCKS IN AUTONOMOUS VEHICLES

Abstract

The development of reliable autonomous vehicle (AV) software architectures depends heavily on the robustness of deep learning frameworks, specifically Convolutional Neural Networks (CNNs). While CNN techniques have fundamentally advanced environmental perception, object detection, and end-to-end steering control, a unified structural consensus balancing mathematical performance with real-world infrastructure constraints remains fragmented. This paper presents a systematic survey of CNN techniques in autonomous vehicle architectures, using a rigorous mixed-methods approach to synthesize the state of the art. The quantitative stream evaluates data from over 120 foundational papers published up to 2026, mapping benchmarks across critical performance indicators, including classification accuracy, parameter efficiency, data loading rates, and inference latency. This stream isolates the performance boundaries of voxelization-based 3D object-detection CNNs against Bird’s-Eye-View (BEV) vision transformers and multi-scale hybrid feature networks. Concurrently, the qualitative stream utilizes thematic analysis and expert structural audits to evaluate the institutional, regulatory, and physical challenges of deploying these models. It focuses on the explainability of "black-box" neural layers under safety standards like ISO 26262, adversarial vulnerability vectors, and sensor degradation under extreme environmental noise or unstructured traffic conditions. By integrating these mixed-method streams through a multi-criteria decision framework, this survey charts the operational landscape and identifies hybrid modular-E2E CNN designs as the current Pareto-optimal deployment paradigm. Ultimately, this work provides a unified, full-stack reference point and an implementation blueprint for researchers, engineers, and policymakers working to integrate automated mobility safely into future smart city transit networks.