๋ณธ ๊ธ€์€ 2020-2ํ•™๊ธฐ โ€œ์ปดํ“จํ„ฐ ๋น„์ „โ€ ์ˆ˜์—…์„ ๋“ฃ๊ณ , ์Šค์Šค๋กœ ํ•™์Šตํ•˜๋ฉด์„œ ๊ฐœ์ธ์ ์ธ ์šฉ๋„๋กœ ์ •๋ฆฌํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ง€์ ์€ ์–ธ์ œ๋‚˜ ํ™˜์˜์ž…๋‹ˆ๋‹ค :)

2 minute read

๋ณธ ๊ธ€์€ 2020-2ํ•™๊ธฐ โ€œ์ปดํ“จํ„ฐ ๋น„์ „โ€ ์ˆ˜์—…์„ ๋“ฃ๊ณ , ์Šค์Šค๋กœ ํ•™์Šตํ•˜๋ฉด์„œ ๊ฐœ์ธ์ ์ธ ์šฉ๋„๋กœ ์ •๋ฆฌํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ง€์ ์€ ์–ธ์ œ๋‚˜ ํ™˜์˜์ž…๋‹ˆ๋‹ค :)


Particle Filtering

Definition.

Tool for tracking the state of a dynamic system(= change in time) modeled by a Bayesian network.
if you have a model of how the system changes in time, possibly in response to inputs, and a model of what observations you should see in particular states, you can use particle filters to track your belief state.


์šฐ๋ฆฌ๋Š” Particle Filter๋ฅผ BBox prediction์— ์‚ฌ์šฉํ•  ๊ฒƒ์ด๋‹ค. ๊ทธ๋ž˜์„œ

"Particle" = BBox

๊ฐ€ ๋œ๋‹ค.


Bayes Filter

Particle Filter๋Š” ์‚ฌ์‹ค Bayes Filter์˜ ์ผ์ข…์ด๋‹ค.

Definition.

Used for estimating the state of a dynamical system from sensor measurements.
"Bayes Filter" works under the process of Predict/update cycle.

Bayes Filter์˜ ์˜ˆ๋กœ๋Š” ์ด ํฌ์ŠคํŠธ์—์„œ ์‚ดํŽด๋ณด๋Š” Particle Filter์™€ Kalman Filter๊ฐ€ ์žˆ๋‹ค.



์•„๋ž˜๋Š” Particle Filter์˜ ๊ณผ์ •์„ ๋„์‹ํ™”ํ•œ ๊ฒƒ์ด๋‹ค.

์ƒ˜ํ”Œ๋งํ•œ BBox๊ฐ€ ์–ผ๋งˆ๋‚˜ Ground truth์™€ ๋น„์Šทํ•œ์ง€์— ๋”ฐ๋ผ weight๋ฅผ ๋‹ค๋ฅด๊ฒŒ ๋ถ€์—ฌํ•œ๋‹ค.

์œ„ ๊ทธ๋ฆผ์—์„œ ๊ฒ€์€์ƒ‰ ์›๋“ค์€ Particle์„ ์˜๋ฏธํ•œ๋‹ค. ์ด Particle์—๋Š” ์œ„์น˜ ๋ฐ์ดํ„ฐ (x, y)์™€ weight w์— ๋Œ€ํ•œ ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ๋‹ค. weight ๊ฐ’์ด ํด์ˆ˜๋ก ์›์˜ ํฌ๊ธฐ๊ฐ€ ํฌ๊ฒŒ ํ‘œํ˜„๋˜์–ด ์žˆ๋‹ค.

(1) Resampling

weight๋ฅผ ์ด์šฉํ•ด particle์„ weighted sampling ํ•œ๋‹ค.

์ด ๊ณผ์ •์„ ๊ฑฐ์น˜๋ฉด, ๊ฐ Particle์ด $1/N$์˜ weight๋ฅผ Uniformํ•˜๊ฒŒ ๊ฐ€์ง€๊ฒŒ ๋œ๋‹ค.

(2) Prediction

resamplingํ•ด์„œ ์–ป์€ ๊ฒฐ๊ณผ๋ฅผ ๋žœ๋คํ•˜๊ฒŒ ํฉ๋œจ๋ฆฌ๋Š” ๊ณผ์ •์ด๋‹ค.

state transition $p(x_t \mid x_{t-1})$์„ ์ ์šฉํ•˜๋Š” ๋ถ€๋ถ„์ด๋‹ค.

์นผ๋งŒ ํ•„ํ„ฐ์˜ ๊ฒฝ์šฐ, ์—ฌ๊ธฐ์—์„œ Linear model assumption์„ ์ ์šฉํ•œ๋‹ค.

\[p(x_t \mid x_{t-1}) = Ax^{t-1} + b + \epsilon \quad ( \epsilon \sim N(0, \Sigma))\]

ํ•˜์ง€๋งŒ, Particle Filtering์€ Sequential Bayesian Modeling์— ์˜ํ•ด ์ง„ํ–‰ํ•œ๋‹ค๊ณ  ํ•œ๋‹ค. (์•„๋งˆ๋„?) ์•„์ง ์ด ๋ถ€๋ถ„์ด ์ž˜ ์™€๋‹ฟ์ง€ ์•Š๋Š”๋‹ค ใ… ใ… 

(3) Measures

predictionํ•œ particle์ธ BBox๊ฐ€ ์‹ค์ œ G.T.์™€ ์–ผ๋งˆ๋‚˜ ๋น„์Šทํ•œ์ง€ Similarity๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ถ€๋ถ„์ด๋‹ค.

(4) Update

Similarity๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ weight๋ฅผ ์—…๋ฐ์ดํŠธํ•ด์ค€๋‹ค.


์ฐธ๊ณ ์ž๋ฃŒ