|
|
Detection and Classification of Vehicles for Embedded Platforms
Skaloš, Patrik ; Hradiš, Michal (oponent) ; Špaňhel, Jakub (vedoucí práce)
This paper evaluates the performance trade-offs of state-of-the-art YOLOv8 object detectors for vehicle detection in surveillance-type images on embedded and low-performance devices. YOLOv8 models of varying sizes, including one with the lightweight MobileNetV2 backbone and YOLOv8-femto with fewer than \num{60000} parameters, were benchmarked across six devices, including three NVIDIA Jetson embedded platforms and the low-performance Raspberry Pi 4B. Various factors influencing performance were considered, such as weight quantization, input resolutions, inference backends, and batch sizes during inference. This study provides valuable insights into the development and deployment of vehicle detectors on a diverse range of devices, including low-performance CPUs and specialized embedded platforms.
|
host ::
RSS.