A Systematic Study of Autonomous Driving Perception: Models, Co-operation and Constraints

As the research on self-driving systems and autonomous vehicles shift from rule-based autonomy to AIdriven autonomy, machine-learning (ML) inference is used increasingly to support perception, prediction and planning with strict timing and safety constraints. Though the accuracy and complexity of the detection/ fusion models have increased rapidly, the commercialization and robustness of these models are limited by where the inference happens. Factors like latency, reliability, energy use, safety rating and communication overhead decide whether the system is commercially deployable. This paper focuses on the inference and deployment of autonomous systems among three regimes: (i) on-vehicle inference, (ii) co-operative edge inference which uses other vehicles and surrounding smart infrastructure to enhance the AV's worldview, and (iii) cloud/ hybrid inference which offloads heavy computation to a remote server. The trade-offs in safety, communication overhead and robustness under operational stress (sensor failure, network overload, heating issues) are studied. The bottlenecks and performance gains introduced by each method are analyzed and open problems and future focus areas for safety-critical autonomous systems are outlined.