ESP32 is now being used for AI workloads such as face detection with camera boards like ESP32-CAM, or the upcoming ESP-EYE board from Espressif Systems themselves, combined with ESP-WHO face detection and recognition framework.
The two aforementioned board require you to use a phone to vizualize the results, unless you blink some LEDs or connect your own display. But the just released TTGO T-Camera board includes a 128×64 OLED display which should allow you to display face detection and/or recognition results, as well as a BME280 environmental sensor, a PIR sensor, and an optional fisheye lens.
TTGO T-Camera board specifications:
- ESP32-WROVER-B Wireless Module
- SoC – ESP32 dual core Tensilica LX6 processor
- Memory – 8MB PSRAM
- Storage – 4MB SPI flash
- Connectivity – 2.4 GHz 802.11n WiFi 4, Bluetooth 4.2 LE
- Camera – 2MP OV2640 camera with normal or fisheye lens
- Display – 0.96″ 128×64 OLED display connected via SSD1306 I2C display controller
- Audio – Microphone
- USB – 1x micro USB port for power and programming (CP2104)
- Sensors – AS312 PIR, BME280 pressure, humidity and temperature I2C sensor
- Expansion – 5-pin I2C header with 5V,3.3V, and GND
- Misc – Reset and user (IO34) buttons, LED
- Power Supply – 5V via micro USB port, support for LiPo battery via 2-pin header and IP5306 charging chip
- Dimensions – 68 x 28 mm
The board is made by LilyGO, and the company just uploaded Arduino sample code to esp32-camera-bme280 Github repository.
TTGO T-Camera can be purchased on Aliexpress with a normal or fisheye lense for respectively $15.58 or $17.22 plus a small shipping fee.
Thanks to Robin for the tip.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
Interesting, but again would perhaps benefit having micro-SD card expansion to supplement the 4MB flash. and allow for things like video/audio recording, image recognition etc. The ESP32 already has the SPI bus for it.
I think the ESP32 is really the wrong chip for this. The Allwinner V3S or S3 would be much better and not that much more expensive.
The ESP32 can run these CNNs (convolutional neural networks), but it is going to take a while. Depending on the size of the CNN these networks will take somewhere between 2 and 20+ seconds to run.
Compare that to the Allwinner V5. The V5 has face recognition hardware and it can give you an answer in 20ms – about 1000x faster. The V5 is $8 vs $2 for the ESP.
I’ve also tried doing this on the V3. V3 runs the CNNs in 0.2 to 5 seconds – depending on the network.
An interesting note, Tensillica makes a dedicated hardware CNN engine. Maybe a future Espressif product?
https://ip.cadence.com/news/564/330/Cadence-Announces-New-Tensilica-Vision-P6-DSP-Targeting-Embedded-Neural-Network-Applications
You can play with that engine right now, it is in the Mediatek Helio P60.
And how does one get p90 then?
It’s easy to get the Helio P60, just buy a phone using it. $130
https://www.aliexpress.com/item/Nokia-X5-Smartphone-3060mAh-3G-RAM-32G-ROM-3-Camera-Dual-Sim-Cards-Fingerprint-ID-5/32953181305.html
The Helio P60 has four copies of the Tensilica engine.
Thank you. Apparently I had been dreaming about p90 ever since it was announced 😉 hence the typo. But p60 will do in the meantime 🙂
Here are some benchmarks
http://ai-benchmark.com/ranking.html
I’ve got the board running without any difficulty, but was surprised to see that the temperature reported by the onboard BME280 is 40degC while the ambient temperature is25degC (measured on another BME280 which is 50cm from this board). The ESP-EYE BME is well away from the WROVER chip, so I would not have expected that much interference. It is as if the ESP built-in sensor is providing the data source, because I’ve run other ESP32 boards, and the built-in temp sensor reads around 40degC.
I can confirm the problem with the temperature. The BME280 is at a “warm” place on the PCB. So I can’t use it as a temperature control ;-(
This won’t be suitable for all projects, but what if ESP32 is set to deep sleep most of the time, and only wakes up periodically to check temperature? I’d assume the PCB may not get as warm, if all all.
Explanation and example code for ESP8266 @ https://openhomeautomation.net/esp8266-battery
I also believe the high board temperature is affecting the humidity reading, as the displayed humidity on the ESP-EYE is 20.39% compared to my ESP32 with ext BME280 which reports 56%, closely matching my commercial weather station (61%), and DHT22 (52%)
The barometric pressure reading is correct, but the temp and humidity are not usable in this board configuration.