The core concept behind the architecture is depthwise separable convolution. The Visual Wake Word model used for our elevator use case is based on an existing MobileNets model. Visual Wake Word (VWW) Model for Person Detection Further post-training quantization methods were implemented to reduce the size of the model (KWS model: 26,720 bytes VWW model: 3,00,568 bytes Total size after compilation: 207712 bytes 79% of dynamic memory) and latency of the inferences in real time. As our development board of choice - Arduino Nano 33 BLE Sense, has limited resources available on board (1MB CPU Flash Memory, 256KB SRAM), the neural network on the system had to be memory and energy efficient compared to a whole scale neural network model. This allows for fast inferences and efficient memory allocation when and wherever required. In such a structure, only one of the two models runs inference on the input data at a time. The KWS and VWW models are combined in a multi-tenant architecture. Person recognition (VWW Detection) and voice recognition (KWS) to operate the elevator.įire safety mechanism inside the elevator to drop installed oxygen masks and signal the main elevator interface.įire safety and detection system outside the elevator on different floors bundled with the features mentioned in point (1).Ī generic block diagram of the entire system is shown below.Ĭontactless Operation using KWS and VWW CNN Models The project is a three-part system encompassing the following subsystems: The overall flow of the system is shown in the figure below. The elevator then operates as it should and the doors should open for him. Then, the person would speak the floor number he would like to visit, and the elevator system recognizes the floor number which was spoken. High-Level Process Flowįirst, we would want a person to come and stand in front of the elevator, indicating that he wants to use it. Lastly, we built a PCB to house all these components and make a holistically deployable product. These systems are deployed outside and inside an elevator, respectively. Further, we developed a fire safety system using basic yet functional sensors in two methods – using a microcontroller and using an operational amplifier. To create an effective pipeline, we used multiple signal processing techniques to extract necessary features from the input signals.

It consists of two parts – person detection and speech recognition. Using an approach of Embedded Machine Learning, we built a multi-tenant architecture to facilitate a contactless elevator system.

January 2022: Check out the Arduino Team's Blog Post covering this project! June 2021: Awarded 'Best Project' out of a pool of 1000+ projects at the 5th IEEE National Level Project Competition. Smart Elevator Systems using Embedded Machine Learning and Fire Safety MechanismsĪpril - June 2021 | Micro Project in SY BTech Updates

Elevator project arduino uno