The ESP32 cam is a low-cost module with both WiFi and Bluetooth capabilities.
We use a pre-trained Yolov3-tiny model for real-time object detection: a compressed version of Yolov3 to improve the model's efficiency and reduce its memory footprint.
The esp32cam Arduino library provides an easy-to-use API with a web server that displays video streams captured by the camera.
The hardware we used is linked here and includes the chip, camera, FTDI USB to TTL Serial Converter, and jumper wires.
Follow the instructions here instructions on hardware pinout and how to run the Camera WebServer.
Clone repository:
git clone vastnet
Run real-time inference.
To view argument options add the --help flag.
python3 inference.py --model_name=<model_name> ...
- Model config files are stored as
darknet/cfg/<model_name>.cfg - Pre-trained weights files are stored as
darknet/cfg/<model_name>.weights - Class names are stored in
darknet/data/coco.names
| Model | Inference time |
|---|---|
| yolov3 | ~400-600ms |
| yolov3-tiny | ~40-60ms |
####Problem with Arduino Setup
-
Be sure that the Arduino configurations are as follows:
- Board: ESP32 Wrover Module
- Upload speed: 15200
- Flash frequency: 40mhz
- Flash mode: QIO
- Partion scheme: Huge App
- Uncomment the line
#define CAMERA_MODEL_AI_THINKERinCameraWebServer.ino
-
Make sure you are using a static IP and entering a valid ssid and password. Ensure that there is no firewall.