As we don't know anything about original thesis You are to write recommendations for, it's not possible to advice what to recommend as an improvement over existing one. If You post a link to the original thesis, someone might want to read it and post what they think can be improved, although I wouldn't count on it, as analysing thesis is Your job, as You'll be the one to get rewarded for the work You do, hence not many people might be willing to do the job for You.
As for "where to start", read the original thesis, make sure You fully understand it, think about requirements raised by Your professor, brainstorm and write down ideas that arise, evaluate feasibility of ideas written down (are skills that You have are enough to to complete it, do You have enough time to complete it, do You have/can You get required resources to complete it). After all that You should have only few/one feasible idea/s making final choice easier.
Implementation phase comes after You have fully decided upon what You want to achieve and requirements raised by Your professor. Going with simulation/mathematical model only might be enough to satisfy Your educator and is going to be way less expensive than building physical robot.
To implement motion capture using stereo-vision You should be able to understand how computer vision works and how to implement algorithms required to track human body and decode limb joint angles. There is vary good C++/Java/Python library for that called OpenCV
To implement simulation, one has to be quite good in OpenGL, or DirectX, or XNA, or any other GPU language to create graphics (a human model/skeleton) to mimic whatever is detected by motion capture system. For starters I wouldn't bother with full blown human, something like skeleton (image attached below) would suffice. I must add that there are software packages for human simulation like the ones from Siemens
. You university might have development access to these, hence You would not need to reinvent the bicycle. I know, that these packages are capable of taking input from motion capture systems and applying it to virtual manikins.
To implement actual robot one is required to understand and have know-how of all of the above plus mechatronics: materials, physics, actuation and control systems. Again, You university might be able to provide You with a humanoid robot, so You don't have to reinvent the bicycle and focus fully on human mimicking
Must add, that if You think of creating whole thing (motion capture system and robot) from a scratch, then You should reconsider, as each of the systems (and even sub-parts of them) is a serious thesis on its own requiring a lot of resources (time, money, knowledge and skills).