Name
Abstract | ||
|---|---|---|
In this paper, we present JanusNet, an efficient CNN model that can perform online background subtraction and robustly detect moving targets using resource-constrained computational hardware on-board unmanned aerial vehicles (UAVs). Most of the existing work on background subtraction either assume that the camera is stationary or make limiting assumptions about the motion of the camera, the structure of the scene under observation, or the apparent motion of the background in video. JanusNet does not have these limitations and therefore, is applicable to a variety of UAV applications. JanusNet learns to extract and combine motion and appearance features to separate background and foreground to generate accurate pixel-wise masks of the moving objects. The network is trained using a simulated video dataset (generated using Unreal Engine 4) with ground-truth labels. Results on UCF Aerial and Kaggle Drone videos datasets show that the learned model transfers well to real UAV videos and can robustly detect moving targets in a wide variety of scenarios. Moreover, experiments on CDNet dataset demonstrate that even without explicitly assuming that the camera is stationary, the performance of JanusNet is comparable to traditional background subtraction methods. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1109/ICCVW54120.2021.00436 | 2021 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS (ICCVW 2021) |
Keywords | DocType | Volume |
Background subtraction, foreground segmentation, moving objects detection, optical flow, UAV, neural network, CNN, video surveillance, tracking | Conference | 2021 |
Issue | ISSN | Citations |
1 | 2473-9936 | 0 |
PageRank | References | Authors |
0.34 | 19 | 4 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Yuxiang Zhao | 1 | 0 | 0.34 |
| Khurram Shafique | 2 | 0 | 0.68 |
| Zeeshan Rasheed | 3 | 0 | 0.68 |
| Maoxu Li | 4 | 0 | 0.68 |