Fast Object Detection Using a Frame Skip Method

The object detection deep neural network such as the YOLOv3 shows great performance in accuracy. However, it requires high computational complexity. This work proposes a method for reducing the overall complexity of the object detection on a video stream. For each frame of the input video, temporal subtraction and ORB feature matching is performed between the two consecutive frames to determine whether to perform the YOLOv3 or keep the detection result of the previous frame. The average operation time of the proposed method is reduced to 91.78% on GPU and 59.38% on CPU compared to the YOLOv3 performed on every frame.