Study Benchmarks Object Detection Models Across Edge Devices for Real-Time AI Apps

A comprehensive benchmark study examines how modern object detection models perform when deployed on constrained edge devices, addressing a critical gap in understanding their practical viability for autonomous systems, surveillance, and smart city applications.

Study Design and Scope#

Researchers tested three families of detection models—YOLOv8 (Nano, Small, Medium variants), EfficientDet Lite (Lite0, Lite1, Lite2), and SSD (MobileNet V1, MobileDet)—across a range of hardware platforms spanning Raspberry Pi 3, 4, and 5 units, some with Coral TPU accelerators or AI HAT+ expansion, plus Jetson Nano and Jetson Orin Nano devices. The evaluation focused on energy consumption, inference speed, detection accuracy, and how performance shifts as the number of objects in a scene increases.

Key Performance Trade-offs#

The findings highlight fundamental tensions in model selection. SSD MobileNet V1 delivered the fastest inference and lowest power draw but struggled with accuracy. By contrast, YOLOv8 Medium achieved superior detection precision yet demanded significantly more computational resources. Orin Nano emerged as the most balanced platform overall, maintaining competitive performance across most model types.

Hardware accelerators—specifically TPU chips—boosted efficiency for SSD and EfficientDet Lite models but paradoxically reduced YOLOv8's accuracy, suggesting that optimization strategies differ across architectures.

Scene Complexity and Model Behavior#

An unexpected insight emerged from analyzing how models handled varying object counts. When images contained few objects, different models produced similar accuracy levels. However, as scene complexity increased and more objects appeared in frames, accuracy gaps between models widened considerably. This suggests that model selection becomes increasingly critical for challenging real-world scenarios with dense crowds or cluttered environments.