Deploying deep learning-based object detection models on low-power edge AI SoCs presents unique challenges, as real-world performance is influenced by more than just theoretical compute power. This study conducts a comprehensive benchmarking analysis to examine the scalability, efficiency, and real-time feasibility of object detection models in constrained edge environments. Our findings reveal that inference latency does not scale predictably with model complexity, as factors such as memory bandwidth, multi-core scheduling inefficiencies, and data movement bottlenecks significantly impact performance. Moreover, multi-core acceleration does not always yield proportional speedups, highlighting the need for better workload distribution strategies. While low-power SoCs show potential for real-time AI applications, achieving efficient deployment requires a holistic optimization approach that balances model architecture, hardware utilization, and inference acceleration techniques. These insights contribute to the broader understanding of deep learning deployment on embedded AI hardware and highlight key directions for future advancements in edge AI optimization.