Satellite Video Object Detection Network based on DLA-34
Jiangtao Meng, Wei Hou, Baoyu Ge, Jinpeng Wang, Nan Su, Yiming Yan · 2023
Now, satellite video data faces challenges such as complex backgrounds, limited spatial features of targets, and artifacts caused by satellite motion, all of which increase the difficulty of target detection. We have designed a Satellite Video Object Detection Network (SVODNet). The algorithm consists primarily of a temporal dynamic feature extraction module and a grouped channel attention module. The temporal dynamic feature extraction module captures the changes of targets over time by utilizing both global and local information, effectively overcoming the influences from complex backgrounds, limited spatial features, and artifacts. Additionally, to further analyze the dynamic changes of targets throughout the entire video sequence, we employ a grouped channel attention mechanism to integrate global and local information. The SVODNet has been evaluated on Jilin-1 satellite data with competitive results.