Real-Time Object Detection in Intelligent Surveillance Videos Based on Improved YOLOv8

Real-Time object detection in the surveillance video stream has very important value to guarantee public safety and make intelligent monitoring practical. To improve the effect of detection in multitask scenarios in terms of accuracy and speed, this paper proposed an improved model based on YOLOv8-MSA-YOLOv8 (Multiscale Attention YOLOv8). The model, based on the original architecture of YOLOv8, is augmented with an added multi-scale adaptive attention fusion module (MAF), dual-path feature enhancement network (DP-FEN) and a lightweight temporal transformer encoder (TTA), which further enhances the expression ability and timing stability of multi-scale targets. The MAF module is added directly between the backbone network of YOLOv8 and the neck structure to optimally capture cross-scale semantic dependencies. DP-FEN jointly enhances shallow texture information and deep semantic information independently, and improves the recognition ability of small targets and occlusion targets. The TTA module, while not sacrificing the inference speed, introduces an inter-frame sharing weight mechanism directing the learning of the target’s motion mode, and further enhancing its modelling ability. The modules are also jointly trained using a mixed loss function composed of an IoU perceived loss and attention- guided regularization. The experimental results showed the MSA-YOLOv8, when compared to the original YOLOv8 model through testing, yielded a $4.7 \%$ increase in the mAP and a $16 \%$ decrease in the false positive rate per frame.