Benchmarking Object Detection Models on Embedded GPUs


Following an ongoing project, we aim to improve upon an existing object detection model for an autonomous drone control system. Furthermore, we aim benchmark the performance of these models on the embedded hardware to choose an optimal model for the given platform. Ultimately, we chose the Yolov4-tiny architecture trained to recognize three types of actors: deer, cars, and pedestrians. This model had a comparatively small memory footprint with very little tradeoff in accuracy and speed.

What We Did

Created benchmark data for Yolov4, Yolov4-tiny, and MobileNet object detection models on the Jetson AGX Xavier. For the model chosen: Yolov4-tiny with 3 classes

  • 91% mAP
  • 95 max FPS
  • 1.2 MB memory footprint

To evaluate an energy conscious GPU scheduler, we need a sufficiently difficult task to evaluate, hence this project. Furthermore, the current object detection models can only recognize deer and cars, as well as misclassifies benches as deer. As such, we aim to improve the number of actors being recognized as well as improve accuracy.

Project Advisors

Client: Oren Bell, Graduate Researcher

Advisor: Dr. Silvia Zhang, WUSTL ESE Professor