This repository contains part of the implementation code for paper ImgSensingNet: UAV Vision Guided Aerial-Ground Air Quality Sensing System (INFOCOM 2019).
ImgSensingNet is a vision guided aerial-ground sensing system, for fine-grained air quality monitoring and forecasting using the fusion of haze images taken by the unmanned-aerial-vehicle (UAV) and the AQI data collected by an on-ground wireless sensor network (WSN). This code contains the vision part implementation of ImgSensingNet, where we realize the end-to-end inference from the taken hazy images to the corresponding AQI values. Specifically, we propose:
- Six content-nonspecific haze-relevant features for raw haze images, which extract the haze components from origin images while eliminating the influence of image content;
- A 3D CNN model for end-to-end learning for AQI estimation of haze images based on the extracted haze components.
The code was tested on
python == 3.6.5
opencv == 3.4.1
scipy == 0.19.1
We propose a list of haze-relevant statistical features. An example of the extracted features is shown above.
The dark channel is an informative feature for haze detection, defined as the minimum of all pixel colors in a local patch.
To obtain a better estimation of haze density, we propose the refined dark channel by applying the guided filter on the estimated medium transmission, to capture the sharp edge discontinuous and outline the haze profile. (dark_channel.py)
The contrast of the haze image can be highly reduced when haze exists. We therefore propose max local contrast feature as the local maximum of local contrast values in a local patch. (max_local_contrast.py)
Image saturation varies sharply with the change of haze in the scene. We therefore define max local saturation feature that represents the maximum saturation value of pixels within a local patch. (max_local_saturation.py)
The scene depth is found to be positively correlated with the difference between the image brightness and the image saturation, which is regarded as the color attenuation prior. To process the raw depth map for better representation of the haze influence, we define the min local color attenuation feature by considering the minimum pixel-wise depth within a local patch. (min_local_clr_attenuation.py)
The hue disparity between the original image and its semi-inverse image (defined as the max value between original image and its inverse) can serve as another haze-relevant feature. (hue_disparity.py)
In the CIELab color space, the chroma is one of the most representative image feature to describe the color degradation by the haze in the atmosphere, which we utilize as another feature. (chroma.py)
feature_extract.py: Generate the six proposed feature images (grayscale, resized to 128*128 pixels) corresponding to a set of input haze images. Run with
python feature_extract.py --data-path <path> --save-path <path>video_sampling.py: Generate haze images with a constant sampling rate when the source file is video (which is more efficient for data collection). Run with
python video_sampling.py --data-path <path> --save-path <path>Note that the extracted feature maps of each image will first be stacked together as an input tensor (which we call as the “hardwired” layer), and then feed into a 3D CNN model for training. More details can be found in our paper.
A real-time demo of the UAV vision-based sensing (including feature extraction and real-time prediction) can be found here.
If you find the idea or code useful for your research, please cite our paper:
@inproceedings{yang2019imgsensingnet,
title={ImgSensingNet: UAV Vision Guided Aerial-Ground Air Quality Sensing System},
author={Yang, Yuzhe and Hu, Zhiwen and Bian, Kaigui and Song, Lingyang},
booktitle={Proceedings of the IEEE International Conference on Computer Communications (INFOCOM)},
pages={1--9},
year={2019},
organization={IEEE}
}