Evaluating the Impact of Computational Perception Architectures on Autonomous Vehicle Safety Through Hardware in the Loop Analysis of Processing Latency andStopping Distance

Abstract

The rapid integration of computer engineering into modern vehicles has shifted automotive safety from predominantly passive mechanical protection to intelligent, software-driven, actively responsive systems. This study investigates the influence of computational perception architecture on autonomous vehicle safety, with particular emphasis on processing latency and its effect on emergency stopping performance. A hardware-in-the-loop (HIL) experimental framework was developed using the CARLA simulator and an NVIDIA Jetson AGX Orin embedded computing platform to evaluate three perception architectures: Vision-Only (VO), LiDAR-Only (LO), and Sensor Fusion (VO+LO). Each architecture was evaluated under identical forward-collision conditions at an initial vehicle speed of 60 km/h, with 100 simulation runs conducted per configuration. The results demonstrate a direct relationship between computational latency, reaction distance, and total stopping distance. The Vision-Only architecture exhibited the highest mean processing latency of 284.17 ms, corresponding to a reaction distance of 4.74 m and a total stopping distance of 23.63 m. In contrast, the LiDAR-Only architecture achieved the lowest latency of 141.05 ms, reducing the reaction distance to 2.35 m and the total stopping distance to 21.24 m. The Sensor Fusion architecture produced an intermediate latency of 197.62 ms and a stopping distance of 22.18 m while providing greater perceptual redundancy and environmental robustness. Overall, the findings demonstrate that reducing computational latency can directly improve emergency braking performance by increasing the available physical safety margin. Although LiDAR-Only provides the fastest computational response, Sensor Fusion offers a more balanced solution between response speed, detection robustness, and redundancy. The study therefore highlights the importance of deterministic, redundant, and computationally efficient architectures for the development of reliable safety-critical autonomous driving systems.