AI Based 5G IP Multicast Optimization Framework for Automotive Systems

This paper presents an AI-driven framework for embedded automotive systems with enhanced 5G IP multicast, secure tunneling, and advanced video processing. The system integrates multi-camera vehicle input across four virtual machines, transmitting video via GRE/VxLAN tunnels to an STM32-based edge device. STM32Cube.AI enables real-time, memory-efficient Multi-Modal object detection and frame optimization using a custom Fast RCNN-FPN architecture with selective pruning. This reduces video size and volume while delivering low-latency insights. Optimized video is wirelessly sent to a hybrid AWS-Azure cloud infrastructure, supporting auto-scaling and cross-cloud redundancy. A centralized Video Management System (VMS) in the cloud enables intelligent analytics and storage optimization. A novel framework analyzes 4G-to-5G network equivalence for network provisioning using data from AWS EC2 instances via AWS IoT Greengrass connected to the STM32 edge. 5G Network Slicing is provisioned using cloud-based reinforcement learning (Proximal Policy Optimization) for predictive caching, achieving 93% network slicing accuracy. Performance results show 92 % object detection accuracy, 27.56 % video size reduction per stream, and network latency of $\mathbf{5} \boldsymbol{- 9} \mathbf{~ m s}$ with 2040 % bandwidth usage and 94.4 % signal uptime. Cloud storage is validated using Azure Blob and Amazon S3. The scope of the proposed unique Embedded AI framework is broad and is modified for embedded automotive applications, offering use cases such as accident analysis, insurance validation, driver behavior monitoring, and predictive safety alerts. For smart cities this proposed system can be added as a feature as it enables real-time road infrastructure monitoring, including pothole and signage detection, with automated alerts to city systems. It also supports remote law enforcement through live incident reporting and evidence streaming.