An improved nighttime people and vehicle detection algorithm based on YOLO v7
Youyu Wu, Yangxintai Tang, Tao Yang
Abstract
To solve the problem that the detection of people and vehicles in the night scenario is inaccurate and the detection speed is slow, which can not meet the needs of the scene, this paper proposes an improved nighttime people and vehicle detection algorithm based on YOLOv7, which aims to give consideration to the detection speed and detection accuracy. The improved algorithm is based on YOLOv7-tiny. In the backbone part, the algorithm adds a Ghostnet V2 module to reduce parameters. The activation function in the convolution layer is also changed from leaklyReLU to FReLU function. And Wise-10 U boundary box loss is introduced to accelerate the convergence speed and increase the detection accuracy; In the head part, the parameter-free attention module SimAM, the C3 module and Omni-Dimensional Dynamic Convolution (ODConv) module are introduced into the model. Finally, according to the dataset experiment, improved algorithm reduces the number of floating-point operations by 59%, the mAP value by 0.64 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> , and the inference speed reaches 47 fps, which shows that the algorithm can have a good performance.