oSATCo App Development:
Automation of video analysis

A placeholder description and if this description had more words

how would that look?

POSE Model

Analysis / Metrics

SATCo Model

Timeseries / Examples

Future Development​

Pose Model (Topology)​

Pose Model (Training)​

  • Leave portion of data for testing
  • Use the rest for training
  • Choose hyperparameters
  • Train until convergence
  • Train cycle: ~16 days (A6000 GPU)
POSE Model Process image
𝑑_𝑖=√((𝑥_𝑖−𝑥_𝑖^′ )^2+(𝑦_𝑖−𝑦_𝑖^′ )^2 )

Pose Analysis (Object Key-point Similarity)

d: Euclidean key-point distance

s: object scale (~0.25 x area)

k: key-point variation (~0.15)

v: key-point visibility

𝐾𝑆𝑖= 𝑒^((−(𝑑_𝑖^2)/(2𝑠^2 𝑘_𝑖^2 )) )
𝑂𝐾𝑆= (∑𝐾𝑆𝑖 𝑣i)/(∑𝑣𝑖 )

OKS: 95.29% > avg. 90.96% < OKS: 86.63%

Pose Analysis (Average Precision)

APt=∑OKS> t

Average Precision at threshold of 50

AP=1/N ∑AP[0.5:0.95] 

Average Precision at different thresholds between 0.5 and 0.95

Pose analysis - Average graph image

Pose Analysis (Results)

Pose analysis - Results statistics image
Pose analysis - Results graph image

SATCo Model (Topology)

SATCo Model Topology flow image

SATCo Analysis (Time-series) – 1

SATCo Analysis Time-series -1 graph image

SATCo Analysis (Time-series) – 2

SATCo Analysis (Time-series) – 3

SATCo Analysis – Results

𝐹1=2𝑇𝑃/(2𝑇𝑃×𝐹𝑃×𝐹𝑁)

Per frame results (relevant segments)

𝑟𝑎𝑛𝑘𝑖𝑗 = 𝑠𝑒𝑔𝑚𝑒𝑛𝑡𝑖 x 𝑡𝑒𝑠𝑡𝑗 x o𝑢𝑡𝑐𝑜𝑚𝑒𝐷

Low SWAP Device and Camera