A Comparison of YOLOv4-Tiny and YOLOv7 Models for Pedestrian Detection

Detection of pedestrians is one of many functions that the computer vision system in an autonomous vehicle has to perform. State-of-the-art object detection models based on deep learning are remarkably accurate but they are large and computationally demanding. There also exist some reduced models that could perform detection faster but are less accurate. In this paper, the trade-off between performance and speed of detection for two models, YOLOv4-tiny and YOLOv7, is studied. Experimental results show that the detection speed of YOLOv4-tiny is about 5.3 times faster than YOLOv7, but its accuracy is only about 67.5% of YOLOv7. By analyzing the network structure of these models, some modifications to YOLOv7 that could improve its detection speed are suggested.