Object detection is a core task in computer vision, and the You Only Look Once (YOLO) family remains a preferred choice for real-time applications. With increasing demand to deploy detectors on resource-constrained devices, many researchers have proposed lightweight YOLO variants. This systematic literature review synthesizes peer-reviewed studies on lightweight YOLO models published from 2016 to 2025. We performed reproducible searches in Scopus and Web of Science, screened records using Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA), and applied a 12-item quality assessment to select 103 peer-reviewed journal and conference articles for detailed analysis. We classify primary lightweighting strategies—pruning, quantization, knowledge distillation, and lightweight network architectures—and evaluate their trade-offs in accuracy, latency, model size, and energy use. We also examine application domains, dataset effects, common evaluation metrics, and deployment platforms ( e.g ., Jetson, Raspberry Pi, microcontrollers, Field-Programmable Gate Arrays (FPGAs)), and highlight practical hardware–software considerations. Emerging directions such as hybrid convolutional neural network (CNN)-Transformer modules and hardware-aware Neural Architecture Search are discussed, and key open challenges are identified, including cross-scene generalization, robustness to dynamic environmental interference, and the need for standardized evaluation protocols and hardware-aware co-design. This review consolidates current knowledge and offers practical guidance for researchers and practitioners developing or deploying lightweight YOLO models on constrained hardware.
Wang et al. (2025) studied this question.