Lightweight Deep Learning Models for Real-Time Object Detection on Edge Devices: A Comparative Performance Study

Abstract

Real-time object detection at the network edge — on cameras, drones, vehicles, and embedded industrial controllers — has become a defining workload of applied computer vision. The governing constraint is no longer accuracy alone but the joint budget of latency, memory, energy, and thermal headroom available on resource-constrained hardware. This paper presents a structured comparative study of the principal lightweight detection approaches, organised around two axes: architectural efficiency strategies (compact backbone design, single-stage detection heads, multi-scale feature aggregation) and post-design compression strategies (quantisation, pruning, and knowledge distillation). We develop a comparison framework covering seven evaluation dimensions — detection paradigm, backbone design strategy, relative model footprint, latency behaviour, accuracy–efficiency trade-off profile, hardware affinity, and deployment tooling maturity — and apply it qualitatively across representative model families including the MobileNet-SSD lineage, the compact YOLO lineage, and compound-scaled detectors of the EfficientDet type. Because published benchmark figures are notoriously sensitive to input resolution, runtime, numeric precision, and target hardware, we deliberately separate the comparative analysis from measurement: the paper specifies a reproducible edge benchmarking protocol — fixed input regimes, warm/cold latency separation, sustained-throughput thermal testing, and energy-per-inference measurement — through which the framework’s qualitative rankings can be instantiated as numbers on any given device. We conclude with deployment guidance matching model classes to application profiles and identify open problems in edge-aware neural architecture search and accuracy-preserving low-bit quantisation.