Unidirectional-Edge Detection based Background Subtraction method for Real-time Object Detection in Restricted Environments
Whui Kim, Ju Derk Park, Jae Hong Ryu, Byeong Cheol Choi, Chang Won Lee · 2023
Although many studies focus on the deep learning algorithms, traditional image processing and machine learning technologies are still being developed and used as an auxiliary means. For examples, there is background subtraction based object localization. This can reduce the number of deep learning model inference. The most famous background method for subtraction is background subtraction using the GMM-derived MOG, KNN, and MOG2 algorithms. However, these algorithms still use a non-trivial of computing resources on lightweight single board computers. To alleviate this problem, we propose a unidirectional edge detection based background subtraction algorithm in restricted environments. In terms of processing time, proposed algorithm outperformed others. Although the processing time improved significantly, the precision (77.8927%) was only about 1% lower than the best method. These improvements will enable the video surveillance system to be implemented on lightweight single board computer, such as NVIDIA Jetson boards.