COUNTING VEHICLES BY DEEP NEURAL NETWORK IN HIGH RESOLUTION SATELLITE IMAGES

摘要

Recently, much more high-resolution satellite images became available. We can detect vehicles in such high-resolution images, and estimation of earnings by counting vehicles in commercial facilities is becoming popular. For this purpose, we need to detect and count vehicles in satellite images accurately. In applications for detecting vehicles, deep neural network has achieved state-of-the-art performance like in general image classification and object detection. To evaluate the accuracy, we tested two methods: Simplified HDNN (SHDNN), which generates sliding windows and classifies them by CNN, and BING-based CNN (BING-CNN), which extract region proposals by BING and classifies them by CNN. In our experiment, while the SHDNN has achieved better performance than the BING-CNN, the BING-CNN was much faster than the SDHNN. And we found some issues to work on for improving the accuracy of them.