Yang, Xuelin (School: The Affiliated High School of South China Normal University)
Since the introduction of dockless shared bikes in 2016, there has been more than 16 million shared bikes in China, and illegal parking has become a serious social issue. However, the common management method of human inspection is ineffective to locate the bikes. The goal of this study is to design a new method for automatically detecting shared bikes in surveillance videos. In state-of-the-art detection networks, advances like Faster R-CNN have reduced the running time, yet there are limitations in processing consecutive images. In this study, I introduce the algorithm of “Faster R-CNN over Attention” (FoA), which can accomplish fine-grained detection of shared bikes in videos. In FoA, an “Attention Region Extraction” (ARE) is introduced to process videos into Attention Regions (defined as video’s background). Then Faster R-CNN would compute the location and brand of shared bikes. An “anchor box optimization” for the network is proposed to generate more specified and robust region proposals with a clustering operation. A dataset with 4,291 images and 12,697 labeled shared bikes is built for the experimentation. The optimized FoA has detection and recognition rates of 94.19% and 92.73% respectively. I generalize FoA into a framework of “Attention Region Extraction plus Region-Based Convolutional Neural Network” (ARE plus R-CNN), discussed as a baseline for specified object detection in videos, such as shopping car, animal, and landscape detection. This study proposes a new management method, expands the utilization of video resource, and provides a possibility to break bottlenecks in video detection.
Third Award of $1,000