Abstract
Abstract Hospital scenarios require stable and intelligent video detection of abnormal targets and personnel behaviors. To improve the detection performance of small abnormal targets in hospital scenarios, reduce false alarms in knife-holding behavior detection, and adapt to the recognition needs of temporal abnormal behaviors such as falls and smoking, this study improves the YOLOv8 network structure by adding a high-resolution feature layer and introducing small target adaptive loss weights. It also incorporates wrist keypoints from human pose estimation to construct spatial position constraints for knife-holding discrimination. Simultaneously, a memory-enhanced two-stream collaborative prediction framework is constructed, fusing appearance reconstruction and optical flow prediction branches to extract spatiotemporal features of behavior. Experimental results show that the improved YOLOv8 achieves an average accuracy of 0.918, a 7.1 percentage point improvement over the original YOLOv8, with an average accuracy of 0.825 for cigarette butt detection and an inference speed of 78.1 FPS. The memory-enhanced two-stream collaborative prediction framework has an area under the curve of 0.928 and an inference speed of 54.1 frames per second, achieving higher average area under the curve on multiple public datasets and maintaining stable detection performance in noisy environments. The proposed detection scheme can achieve real-time detection of abnormal targets and abnormal personnel behavior in hospital scenarios, matching the hardware deployment and application requirements of hospital security scenarios.