I remember in the 1980s the Germans were working on an AI programme to automatically detect an armoured vehicle... a tank... so they got together a huge number of still photos and ran an AI programme over it... the idea being that if it can automatically detect a tank then you can use it to scan a live video feed on a camera on a tank in the field and it could automatically alert the crew to the presence of a tank... enemy or otherwise...
During tests it got to 96% success rate in the lab with still images, so they went to the field test stage and it was terrible its performance was awful... But why?
After years of investigation and testing they realised their mistake... their sample photos with tanks in them were taken on sunny days, and the photos with no tanks were taken on cloudy days and that is what the AI learned to detect if it was a cloudy day or a sunny day... and in that role it was actually rather good.
As you can probably tell the remedy is to change the sample images so sunny day and cloudy day and rainy day and snowy day and even night no longer become a factor... which means very careful selections of sample images... and also images from any environment your vehicle is likely to operate in... like desert or arctic or forest or jungle etc etc.
Modern cameras can easily detect the human face for example and can track it in moving video... and of course Soviet and Russian systems like the tracking systems in Kornet and TOR and Pantsir and Tunguska can also perform such operations too... but it wasn't easy.