I have images from UAV which are geo-tagged. I wish to geo-reference images which were shot in sequence, and want to create a mosaic, which further will be a small map.
Is there any possible way in QGIS to do so?
Assuming you don't want to do this manually by georeferencing each individual image:
No it's not possible to do this with QGIS.
You need specialized photogrammetry software to do this. Something like Pix4D or Agisoft Photoscan (30-day free trial of the full version is available) to name a few or similar. Most of these are very expensive (for individual/hobbyist use) especially if the output needs to be georeferenced orthophoto.
ESRI recently begun offering their own Drone2Map tool which just came out of Beta a few months ago. It uses the Pix4D engine and is really a dumbed down version of Pix4D with a different skin but nicely integrated in the ESRI ecosystem. Unfortunately at ~ $4,000 USD / year it's basically the same price as a full version of Pix4D (Pix4Dmapper Pro) so I am not sure who and why would even consider this particular option.
There are some free open source options such as the VisualSFM, CMVS route but you will need patience, taught skin and a NVIDEA card for the CMVS bit and there is no georeferencing option for exporting an orthophoto. It also hasn't been updated and is brutally slow. Way, way slower than the modern paid options.
The best open source free option and the only one i know of that can output georeferenced orthophotos is the Open Drone Map toolkit. This only runs on linux and is command line only but it works well. It's not as polished and feature rich as the heavy hitters' software. The orthophoto mosaic output is not corrected for exposure (the paid tools do this automatically) which means that if one image is underexposed (darker) and the neighbouring image overexposed (lighter) the transition from one to the other in the resulting orthophoto is not gradual. Orthorectification itself is also a bit harsh not as gradual and seamless as in the paid packages. Most of all, at least in our tests, georeferencing was off by ~ 10 - 15 m. Perhaps some of these can be controlled with some tweaking of parameters which we did not try. We went with all defaults in our tests. And of course you need to run this on Linux which may require