One question people have when they are developing an IoT device with computer vision on IoT device is to know the right camera for the device that works with the Algorithm. This can be a quite complex issue. You may need computer vision on a drone performing analysis of 100s of different types of objects in real time. You may also just need to recognize a face and have a very low price point for your device.
In this article we are going to discuss the hardware classes, camera resolutions, and algorithms need to make your computer vision project successful in IoT.
Device Classes
Not all IoT devices are the same. But there are certain classes of hardware
ESP32
The ESP32 and ESP8266 are low price options. These devices typically pose less than $5. There are several preconfigured options to chose from that can help accelerate your project. The ESP-EYE for example has a camera attached to an ESP32 but with its limited memory and compute is limited to the the to what it can do. When using ESP line you will be very limited to what it can do and it will be very slow. One option that people use to offset this is to transmit the images to the cloud for further analysis.
Price $5 Use Case: Cloud Models, Face Detection
Raspberry Pi Class
Although Raspberry Pis are not hardened for IoT there are many manufactures that do provide great solutions for about the same price point and compute. With a price point of $35 and and using an Linix bases OS it can be a very attritive option. IoT cloud providers such as Azure IoT Hub have built in over the air (OTA) update options as well as its ability to use containers make developing and updating Raspberry Pis class of devices easy to manage.
Price: $35 Use Cases: Smart Home, Object Classification
LottePanda Class
Single Board Computers (SBC) are capable of running a heavy combinations of sensors and cameras. They can run different operating systems and have enough spare compute have monitors and other peripherals.
Price $100 Use Cases: Sentiment analysis Kiosk, high-frequency devices such as heart monitors.
i.MX Series
The i.MX series is a series of SBC that are open source and have a large amount of RAM and CPUs.
Price: $200-$300 Use Case: high speed face recognition, high resolution object detection.
Manifold 2-C with NVIDIA TX2
NVidia makes a series of high end SBCs. They have a echo system of containers that allow teams to onboard on their platform easily. They also have specialty devices such as the TX2 that is made to attach to a DGI drone.
Price: $500 Use Case: Drone object detection, Drone self driving
Specialty Devices
Some IoT devices such as Google’s Coral.AI TPU is only compatible with the Coral Camera. Which limit you to a certain resolution of Camera. I am not really going to discuss those types of devices because they tend to be toy devices that are not hardened to use for IoT use.
Cameras
Getting the camera hardware right can be a very challenging aspect of the equation. First cameras, especially for IoT devices can vary greatly in terms of lens quality, field of range, and image accuracy. The best way to determine the best camera is to take photos with those camera in the same lighting and distance you might expect in field conditions. Some cameras handle low light situations better than others.
Algorithm
The final part of the equation is the algorithm algorithms can be very resource intensive like Yolo or less resource intensive like Haar Cascade. Each of these algorithms has a sweat spot they work best. The best way to find the algorithm, camera, and compute, in the end, to experiment.