Athlete target detection method in dynamic scenario based on nonlinear filtering and YOLOv5

In this paper firefly optimization algorithm is improved, including the method of relative firefly fluorescence brightness, the method of attraction and the method of firefly position update. The dynamic step size factor and dynamic difference factor are introduced, and the improved firefly algorithm is used to optimize the particle filter, so that the particle swarm can be concentrated to the high likelihood region as much as possible, so as to ensure the overall quality of the particle swarm. In this paper, the most commonly used non-maximum suppression algorithm in the post-processing stage of target detection model is discussed. The original YOLOv5 model and the fusion model of nonlinear filtering and YOLOv5 were respectively used to simulate the two data sets after data enhancement. Some mainstream object detection models and the improved model in this paper are analyzed experimentally in two datasets. In dataset 1, the small increase in mAP value is the addition of CBAM's attention module, which increases by 2.8%. For data set 2, when the original Focal loss was replaced by VFLoss, the mAP increased to 89.8%, an increase of 0.95%. For the case where all the improvements were added, the mAP value increased by 7%.