Edge analytics on resource constrained devices

Abstract

Camera sensors can measure our environment at high precision, providing the basis for detecting more complex phenomena in comparison to other sensors, e.g., temperature or humidity. Using benchmarks, this work evaluates object classification on resource constrained devices, focusing on video feeds from IoT cameras. The models that have been used in this research include MobileNetV1, MobileNetV2 and faster R-CNN that can be combined with regression models for precise object localisation. We compare the models by using their accuracy for classifying objects and the demand they impose on the computational resources of a Raspberry Pi. We conclude that the faster R-CNN model that is configured with the InceptionV2 regression model has the highest accuracy. However, this is at the cost of additional computational resources. We found that the best model to use for object detection functionality on the Raspberry Pi is the MobileNetV2 model paired with the SSDLite regression model.