Human Detection with A Feedforward Neural Network for Small Microcontrollers

Abstract

Motion detection is widely used in smart home applications. Nevertheless, it is crucial that a non-human movement does not trigger a motion sensor for various applications. For this purpose, camera-based sensors and machine learning are used. However, modern machine learning, such as deep neural networks (DNN), requires powerful hardware, which is expensive, has high power consumption, and requires a significant amount of storage space. Therefore, DNN object detection methods cannot be used directly on resource-constrained devices as the models require too much storage space. We present an object detection method for a resource-limited system where a camera-based human detection is computed directly on a small microcontroller to overcome these limitations. The key innovation enables object detection on a microcontroller with a small storage space and low computational power. With a unique image pre-processing and a feedforward artificial neural network (ANN), we reached a parameter reduction of the ANN of over 99 % compared to DNN image classification and object detection methods, allowing us to use the algorithm on an ESP32 with a detection rate of 83 ms. Furthermore, we also propose an approach that automatically generates a dataset from videos for the training of the ANN.