Advanced vehicle surveillance: a real-time multi-camera AI-based person activity recognition on a low-power embedded GPU

This thesis investigates the design and implementation of a real-time multi-camera AI-based person activity recognition system on a low-power embedded GPU platform. The work focuses on the feasibility of deploying a complete perception pipeline, including human detection, action classification, visualization, and streaming, under the computational and energy constraints of embedded hardware. Starting from a baseline GPU implementation, the system is progressively redesigned through batching strategies, CUDA kernel adaptations, multi-thread and multi-stream execution, cross-architecture deployment, and embedded-specific optimizations. The final implementation is deployed on the NVIDIA Jetson Xavier NX and validated in a three-camera configuration. The results show that real-time operation is achievable on the selected embedded platform, while also highlighting the practical limits imposed by resource contention, power modes, and model behaviour. More broadly, the thesis shows that embedded AI deployment is not only a model inference problem, but a system-level engineering challenge involving data movement, concurrency, scheduling, and hardware-aware design.