I am trying to look the the practical aspect of using CSI. To this end, I separated the NoActivity data from the given activities and trained the RNN model to classify 8 activities (including NoActivity) and made it work in real time. Although the performance on the dataset in quite good, in practical environment, it does not perform well. The type of environment and disturbances outside the LOS of the transmitter and Receiver also cause the output to change.
I do not know how exactly your 'sitdown', 'standup' or 'fall' activities were performed or how your environment was, but if I try doing these in my room, it can hardly classify them correctly.
I am going to try using directional antennas and changing other aspects (like increasing frequency back to 1kHz and decreasing the number of activities to classify). Can you suggest something else to make it more robust in practical environments?
Hi,
I am trying to look the the practical aspect of using CSI. To this end, I separated the NoActivity data from the given activities and trained the RNN model to classify 8 activities (including NoActivity) and made it work in real time. Although the performance on the dataset in quite good, in practical environment, it does not perform well. The type of environment and disturbances outside the LOS of the transmitter and Receiver also cause the output to change.
I do not know how exactly your 'sitdown', 'standup' or 'fall' activities were performed or how your environment was, but if I try doing these in my room, it can hardly classify them correctly.
I am going to try using directional antennas and changing other aspects (like increasing frequency back to 1kHz and decreasing the number of activities to classify). Can you suggest something else to make it more robust in practical environments?