Accelerate the Detection Frame Rate of YOLO Object Detection Algorithm

Abstract

YOLO (You-Only-Look-Once) is by far the well-known Deep Neural Networks (DNNs) object detection algorithm with real-time performance on a computer with GPUs. Conceptually, YOLO divides the input image of size $W \times W$ into non-overlapping square cells with the final feature of size $S \times S$; i.e. $(416 \times 416) \rightarrow (13 \times 13)$. Each cell is responsible for predicting a single object whose centre falls into it. In this paper, we propose the algorithm that makes use of our observation mapping relationship which states that while the sizes of square cells are changed from layer to layer, their indices are preserved. The algorithm operates by locating a region of change in an input image and identifies the indices of square cells that cover the region. Only the members of the input features within these cells in all layers along the network are required to be operated. When the algorithm is employed along with the spatio-temporal property within video frames, it is capable of attaining the best relative detection of 1.47 (about 7 fps) with 90% correctness. These are benchmarked with the ordinary YOLO object detection on a personal computer: Intel Core i7 CPU at 3.5 GHz with 16 GB of memory and without any sophisticate GPUs, on the Tiny-YOLO network.