Edge Deployment of Multi-Task Vision Models for Smart City Infrastructures

The ability to deploy complex vision pipelines at the edge is crucial for next-generation smart city infrastructures. Running inference locally avoids transmitting raw images, thereby reducing bandwidth usage, preserving privacy, and improving system robustness through a distributed paradigm. Achieving this, however, requires models that are both powerful and efficient. In this context, the multi-task paradigm emerges as a key enabler, supporting concurrent prediction across diverse tasks such as object detection, panoptic segmentation, and depth estimation. We introduce a novel pipeline built around a multi-task model explicitly designed to balance expressive power with inference efficiency under resource-constrained conditions. In parallel, we propose an optimization workflow that minimizes inference costs for edge deployment. Experimental results in a real-world setting demonstrate the effectiveness of our approach and establish a strong baseline for future smart city applications.