Counting Boats and Where They are Heading

I've not come across anything similar myself. I do know there's a fair amount of research published on vision based road traffic vehicle identification and counting systems, which could be modified to suit. There's also a fair amount of tutorials out there that show you how to use OpenCV (Python) for real time vehicle detection and identification, for example. Similarly with TensorFlow.

Cameras and image processing is quite power hungry so would need a decent power source etc. but no doubt a solution can be engineered to suit environment.

As this is a PoC, you could just start with a Single Board Computer like a Raspberry Pi, for example and add a high quality camera - there's also a roadtest on this Raspberry Pi HQ Camera - Review . Alternatively you could try a low cost ESP32CAM, but I suspect this may struggle with this sort of application. Anyway, I'm sure other E14 members will offer advice.

BigG - much appreciated.  Will look into those suggestions.

You could set up a bright LED or laser with a photodetector on either side and a retro reflector across the canal.

Whichever detector loses signal first tells you which direction the boat is going.

The signal probably should be modulated at a high frequency, so the detector signals can filter out sunlight.

A pair of 'break beam' type sensors could potentially give you count and direction.

Some IP CCTV cameras have the ability to set motion detection zones and trigger events upon detection. Could also perhaps be used to help collect sample data sets of images for training machine learning systems later on.

Here's another very recent tutorial on how to get started with object recognition using Edge Impulse and TinyML on a Raspberry Pi.

https://www.raspberrypi.org/blog/edge-impulse-and-tinyml-on-raspberry-pi/

There is also a range of vision related projects using Beaglebone AI on the Element14 website to guide you too: Vision Thing

There is quite an easy going introduction to using OpenCV with Python for object tracking on YouTube:

   Object Tracking with Opencv and Python - YouTube

You perhaps have an advantage that canal boats tend to move a little slower.

It assumes though that you are allowed to set-up a camera securely in an appropriate location which can sometimes result in a lot of paperwork.

Perhaps also consider this if using a R-Pi HQ Camera based project:

   Project14 | Photography: We're Giving Away Raspberry Pi HQ Camera Kits for Projects that Use Them!

Although I don't have any personal experience using this product, I thought for completeness to suggest that you could also the Arduino Portenta H7 (https://store.arduino.cc/portenta-h7 ) and the Arduino Portenta Vision Shield for your application. I note that the vision shield also includes a LoRa module, which may be of use https://store.arduino.cc/portenta-vision-shield-lora.

According to the vision shield description on the webpage, Arduino has teamed up with OpenMV to offer object recognition etc. There's also a tutorial available: https://www.arduino.cc/pro/tutorials/portenta-h7/por-openmv-bt

This short paper on the IOP Science site may be of some interest. It's looking at image recognition of small boats for collision avoidance purposes using Python, TensorFlow and OpenCV.

Use of deep learning algorithms for real-time detection of vessels in confined spaces using the Tensorflow framework - IOPscience

Interesting abstract.  I wonder what they classify as a "small object"?  Boat?  Buoy?  Rock?  If 22% are boat on boat, what are the other 88%?  Boat vs. land?  Great concept.

Another one.

Boat Detection - Image Processing with a Convolutional Neural Network

https://nickgrattandatascience.wordpress.com/2017/09/03/image-processing-with-raspberry-pi-zero/