Animal Crossing WiFi CSI

This dataset is shared as part of the paper Detection and classification of animal crossings on roads using IoT based WiFi sensing, submitted to the IEEE LATINCOM 2023 conference. It is distributed under the Creative Commons license Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). An expanded version of this dataset is available at IEEE Dataport. General description Each sample from the dataset contains 500 frames of WiFi Channel State Information data, captured during a 5-second window (100 Hz sampling rate). Each frame contains the amplitude information from the 52 Wi-Fi subcarriers that transmit a data. This amounts to 26,000 features per sample. The non-zero amplitude values are converted to decibels, while null values are set to zero after the decibel function application to prevent negative infinite values. Subsequently, a running mean filter is applied to each frame to mitigate noise and outlier interference, ensuring a more stable representation of the signal. Additionally, we disregard zero-valued amplitudes, as these result from errors in the original signal capture process, leading to subcarriers without meaningful amplitude. Thus, zero values are not included in the running mean computation. We collected the CSI data using ESP32 boards, which were placed at a height of 70 cm and 12 meters apart from each other. To avoid bias towards a single environment, we collected data in four different locations, including paved and unpaved rural roads, a pasture and a gravel road. The parquet files can be easily read and manipulated with python libraries such as pandas. Data labels As it is intended to allow replication of the work presented, we uploaded the same separated test and training datasets used for the machine learning model. The data labels represent the following classes: 0 - Background noise 1 - Person 2 - Car 3 - Dog 4 - Cow