Vision-based feet detection power liftgate with deep learning on embedded device

Abstract

Kick-Open power liftgates are the standard configuration for most high-end SUVs and sedans in recent years, but the traditional sensor-based trunk opening systems mainly rely on detecting the distance change between foot and vehicle to monitor the users’ operation intention, which is more sensitive to interference like weather change or unintentional body movement. We proposed a novel vision-based feet position detection power liftgate system, which detects the feet’ position through a deep learning model. Its framework is based on Nanodet, and the model is quantified and then convert and deploy on the low-power embedded system. We have trained the network using our dedicated dataset that considered various situations, including various weather conditions, different illumination conditions, different types of shoes in various seasons, and various ground surfaces. The total parameters of our model are 0.7M. In the test phase, our system works at 8FPS and reached a recognition accuracy rate of 93.3%, and a recall rate of 92%. Our system can be applied to low-cost embedded devices for automobiles and show satisfying performance and reliability.