Personalized behavior monitoring for i-secure security applications
- Detailed Technology Description
- We developed 3 software: First - Unsupervised miner of human trajectory from unannotated video. Detect humans in each frame using ACF (Aggregated Channel Features) and then assembling sequences of bounding boxes in image sequences. To predict next coordinate of box Kalman Filter is used. The criteria for assignment of trajectories is quite strict, enforcing that one trajectory belongs to the single human on video, without individualization. Second software - Fit the model to the data, using human-aided annotation (example .xlsx Excel file included) and .h5 tracklets from 1st software. At the moment was tested identification of age, sex, general appearance and luggage, we have identified 8 parameters. Running this software is very time consuming, because it essentially train neural networks, and not single one. It also require strong GPU. On my Nove Hrady PC it takes around 20 min for all. Network themselves are convolutional + LSTM (recursive units with long-short time memory). They are not very complicated by design, but the technologies are cutting edge. Trained model is included.Third software - The software uses the .h5 tracklets from the first software and model from the second software can estimate desired parameters in new sequence of data. Software need instalation of MatlabRuntime 9.7. All algorithms and procedures have been verified and the vast majority of commercial prototype software works. There is also a prototype of connecting the camera to the micro-supercomputer Jetson Xavier
- Application Date
- 19/02/2020 00:00:00
- ID No.
- 296300
- Country/Region
- Czech Republic
For more information, please click Here
