A lightweight dense pedestrian detection and tracking algorithm based on improved YOLOv8 and deep sort

Aiming at the problems that the dense pedestrian tracking accuracy is not high and the tracking speed cannot achieve the real-time specifications in complex scenarios, this paper proposes an improved multi-target pedestrian tracking model, improving the Deep SORT base framework and integrating YOLOv8 to realize the detection and tracking of pedestrians. The CA attentional framework is incorporated into the deep residual framework, which can adaptively select the best convolutional kernel to boost the characteristic extraction ability of the neural framework; the lightweight Ghost is employed to replace Darknet53 to lessen the number of parameters in the model; and the EIOU loss function, which has a higher weight in the localization loss, is utilized to supplant the CIOU loss function of the primary model, to improve the target localization accuracy. Adopting pedestrian re-recognition network to elevate the tracking rate of the Deep Sort model while preserving accuracy. The effect of various improvements on the model performance is verified through ablation studies, and the improved model is analyzed to the current mainstream pedestrian detection tracking models. The experimental outcomes demonstrate that the improved model is effective and advances 9.7% over the MOTA effectiveness of the original algorithm on the MOT16 tracking data repository, and outperforms several other tracking models when compared to them. The algorithmic model is robust for tracking accurately and efficiently even when dense pedestrian movements or pedestrian targets are occluded in complex scenes.