computer vision - How to classify gameplay situations with CNNs (in soccer analytics) -

even though don't have particular code question, try explain story + ask questions precise , specific possible. i'm sort of ml rookie , i've spent time understand basics of computer vision , deep learning little project of mine going. huge soccer/football fanatic , passionate game analytics , game tactics. i've learned more cv , dl in cs class took @ college, thought combination of both might interesting approach (i'm aware not new topic in sports broadcasting).

https://media.giphy.com/media/jl2roao1p3m92/giphy.gif

this gif concrete example of gameplay sequence highly useful game outcome (create chance score goal): player running towards goal, passing player, in doing taking 4 opponents off game , creating (almost) big chance score. type of pass way more useful (valuable) random pass b somewhere within own half left right.

game statistics these days not take consideration quality of action, rather consider quantity of events (like passes in general, possession, duels etc.). wonder if (considered availability of such video footage)

1. it might able classify quality of such gameplay scenarios cnn/rnn example (-> not random pass, pass in final third of pitch, played vertically (i.e. towards opponent goal), taking off x opponent players etc.)
2. if cpu/gpu can process such amounts of data in real-time (considered soccer rather dynamic sport 22 players move @ same time time)
3. if data quality (video footage) above high enough

i've talked phd students this. told me should consider taking intermediate steps (like classifying basic pass direction first). mentioned ltsm , deep reinforcement learning (even though don't have agent i'm not controlling game or simulation; i'm not sure if drl applicable then).

i'd super happy if provide me more hints can more specific research. like: 'yes possible, go , search x or consider y'. or 'no because read abc'.

thanks + sorry bother!

this extension of activity recognition using video. there fair amount of research on it. can have @ topic, , competitions activitynet.

lstm, 3d cnn based architectures , facebook's c3d examples can analyze videos , come sort of results. have tried first 2 , had results datasets virat , kth (basic activities), third big gpu.

my suggestion start simple this, , build on find both data , architectures match task better.