Creating a model from a scratch can be challenging and time consuming. Instead of building a machine learning model from scratch, you can use the starter projects. IMAGIMOB Studio provides a wide range of starter projects that help you to start training and deploying machine learning models quickly.
A starter project allows you to use a machine learning model that is already trained by Imagimob for different scenarios. You can use the starter project as -
- a starting point and fine-tune the model as per your requirement
- an inspiration and collect your own data to build a similar project
- deploy the pre-trained model to an edge device and evaluate the performance
A starter project directory contains the following sub-folders, tools and resources -
- Data folder - contains the sample data
- Models folder - contains the trained model (.h5 Tensorflow model) and edge code, ( neural network model and preprocessor translated to C code) ready for integration into the embedded firmware
- PreprocessorTrack folder - contains the sample predictions from the model
- Units folder - to add custom layers and pre-processors
- Readme.md file - describes the project and data in detail
- Project file - contains the necessary resources to build the machine learning model
The starter projects are intended to be a demonstration of how you can build a machine learning model using IMAGIMOB Studio, different types of sensors and development boards.
Imagimob provides the following pre-trained starter projects -
The following starter project are available for Syntiant and generic development boards.
- Human activity recognition
The human activity recognition starter project provides a machine learning model that predicts human activites such as running, standing, walking, sitting, jumping. The project uses the BMI160 Inertial Measurement Unit (IMU), setup to collect data at 50 Hz using ± 8g for the accelerometer scale and ± 500 dps for the gyroscope scale.
- Keyword spotter
The keyword spotter project provides a machine learning model that recognise keywords like Up and Down with a microphone. The project uses 16kHz mono to record multiple recordings from different people and/or background noise.
- Fall detection
The fall detection starter project provides a machine learning model that detects fall using an Inertial Measurement Unit (IMU) mounted on the buckle of a belt. The project uses Bosh and ST-Microelectronics IMU to collect data at 50 Hz using ± 8g for the accelerometer scale and ± 500 dps for the gyroscope scale.
- Indoor or outdoor detection
The indoor or outdoor detection starter project provides a machine learning model that detects whether a person is indoor or outdoor using the environmental sensors. The project uses the Nordic semiconductor, Nordic Thingy:91 to collect environmental sensor data such as air quality, air pressure, humidity and temperature at different locations in Sweden during the course of two weeks.
The Acconeer radar gesture starter project provides a machine learning model that recognize gestures such as push, wiggle and vertical finger rotation. The project uses the Acconeer A1 radar, which collects data at a sampling rate of 39 Hz. The radar sensor is alow power, high precision, pulsed short-range sensor with a footprint of only 29 mm[^2]. The sensor is configured for a distance of 7-20 cm.
The Texas Instruments radar gesture starter project provides a machine learning model that recognize gestures such hand swipe left to right and right to left, finger rotation clockwise and counter-clockwise. The project uses the Texas Instruments mmWave Radar, IWR6843AOP and AWR6843AOP, which are single chip 60 GHz to 64 GHz radar sensors integrating DSP, MCU and radar accelerator for automotive and industrial applications. The project uses incoming radar data (the cartesian coordinates and speed) of the closest objects to classify different gestures. The data is collected with a sampling rate of 10 Hz.
The Renesas touchpad letter detection starter project provides a machine learning model that recognise letters written on touchpad. The project uses the Renesas RA2L1 board with KT-CAP1-MATRIXPAD as a capacitive touch input. The capacitive touchpad is used to collect data at a rate of 44 Hz.