Machine Vision in Real-Time Environments: A Narrative Review
TL;DR
The literature falls short on consistent real-time benchmarking, robustness/energy reporting, and mature foundation-model integration; the results show real-time performance is achieved through pairing lightweight architectures with hardware-aware compression rather than accuracy gains alone.
Abstract
Machine vision systems that sense, interpret, and act on visual data within strict time limits underpin applications from autonomous driving and industrial inspection to UAVs, precision agriculture, and image-guided surgery. Background: Deep learning has pushed detection, segmentation, and tracking accuracy well past earlier methods, but real-time deployment on power- and memory-limited hardware remains a substantial engineering challenge, and large foundation and vision–language models raise a further question: whether such capabilities fit within a latency-bounded pipeline at all. Methods: This review draws together 94 sources (2012–2026) following a narrative-synthesis design with explicit eligibility criteria, covering detection/segmentation architectures, compression, hardware platforms, application domains, and foundation/vision–language model integration, including compressed large models (e.g., Segment Anything) and quantization, pruning, and distillation on GPU, FPGA, NPU, and microcontroller (TinyML) hardware. Results: Real-time performance is achieved through pairing lightweight architectures with hardware-aware compression rather than accuracy gains alone; trade-offs vary substantially with backbone, resolution, and platform; and foundation/vision–language models remain used mainly as offline data engines or asynchronous reasoning layers. Conclusions: The literature falls short on consistent real-time benchmarking, robustness/energy reporting, and mature foundation-model integration; we close with concrete future directions.