The team began with a practical problem: hospital cameras must recognize targets that differ dramatically in size and behave in ways that cannot be inferred reliably from a single image. A cigarette butt can be extremely small, whereas a person and a stretcher bed produce much larger shapes. Falls depend on rapid changes in posture and position, while smoking can appear as a subtle interaction near the hand and mouth.

Hospitals also introduce practical complications such as changing illumination, compression artifacts, partial occlusion and dense pedestrian traffic. The researchers therefore combined three linked components: a higher-resolution object detector, pose-guided checks for knife handling, and a temporal anomaly model that evaluates both visual appearance and motion. The first component modified the feature pyramid of YOLOv8 by adding a P4 feature-map branch with fourfold downsampling.

This branch preserves finer spatial information than the original layers, which operate at downsampling factors of 8, 16 and 32. The network also assigns scale-aware weights to classification and bounding-box losses, giving small objects a larger share of the training signal. In the reported formulation, a target’s relative size is used to calculate a weight that approaches one as the object becomes smaller; a nonlinear hyperparameter controls the strength of that adjustment.

Binary cross-entropy was retained for classification, while regression continued to use Complete IoU and Distribution Focal Loss. In the ablation experiment, the baseline YOLOv8 produced an mAP@0.5 of 0.742 and a smoke-detection average precision of 0.612. Adding the high-resolution layer raised those values to 0.801 and 0.698, and adaptive weighting increased them further to 0.845 and 0.752.

The system adds a second spatial safeguard for knife handling. Simply finding a knife box overlapping a person box can raise an alarm when someone passes near a stationary knife, so the researchers used YOLOv8-Pose to estimate 17 body keypoints for each person. Left and right wrist positions were selected for verification.

A person and knife must first overlap above the configured bounding-box threshold, after which the Euclidean distance from the knife center to the nearest confident wrist must be less than 0.1 times the smaller image dimension. This location-level rule contributed another 3.3 percentage points to mAP@0.5 in the ablation study. With the complete detector, overall mAP@0.5 reached 0.918 and smoke average precision reached 0.825.

Across five independent runs, overall mAP@0.5 was reported as 0.918 ± 0.009, with a 95% confidence interval of 0.907 to 0.929. For temporal abnormalities, the researchers built a memory-enhanced two-stream collaborative prediction framework. One stream reconstructs the current video frame and is intended to learn the appearance of normal activity.

A second stream predicts optical flow from recent frames and models how movement should evolve. Each stream uses an encoder-decoder architecture with memory modules inserted at multiple points. The modules store feature vectors and retrieve the most relevant entries through cosine similarity, then reconstruct the input feature as a weighted combination of those entries.

An entropy penalty encourages sparse retrieval, while reconstruction, gradient and optical-flow smoothing losses preserve image edges and local movement consistency. Features from corresponding encoder layers in the two streams are fused, allowing appearance changes and motion deviations to complement each other rather than being interpreted independently. The two error signals also determine which type of anomaly is more prominent.

For a fall, a rapid body-axis tilt and displacement are expected to produce a stronger motion abnormality than appearance abnormality. The fall branch therefore requires the normalized motion saliency factor to exceed 0.6 and the fused anomaly score to exceed 0.7. Smoking is more local and subtle, so its branch emphasizes the appearance signal and requires a detected cigarette butt to fall within a person’s bounding box and near a wrist keypoint.

Scores are averaged over consecutive frames, and an alarm is issued only when the smoothed value exceeds the anomaly threshold for a minimum duration. This sequence of checks means the reported fall and smoking detections are not based on motion alone; smoking also requires spatial verification with a small-object detection. Evaluation combined the ShanghaiTech Campus dataset, which contains more than 330,000 frames from 13 scenes, with approximately 150,000 frames from a self-built hospital dataset covering registration halls, wards and corridors.

The hospital data were labeled for falls, smoking, knife handling and crowd gathering, then divided into 70% for training, 15% for validation and 15% for testing. Experiments ran on an Nvidia RTX 3090 GPU with an Intel Core i9-10900K processor. The two-stream framework achieved mean AUCs of 0.928 on the hospital dataset, 0.935 on ShanghaiTech, 0.958 on UCSD Ped2, 0.945 on Avenue and 0.838 on UCF-Crime.

Its strongest reported results were therefore on UCSD Ped2 and Avenue, while its average AUC on the hospital set was lower. Across five independent runs, the hospital AUC was 0.928 ± 0.007 with a 95% confidence interval of 0.919 to 0.937, while the corresponding values were 0.935 ± 0.007 on ShanghaiTech, 0.958 ± 0.005 on UCSD Ped2, 0.945 ± 0.007 on Avenue and 0.838 ± 0.012 on UCF-Crime. The authors note that the current study covers four anomaly types and does not yet establish performance for sudden or complex interactions.

Robustness and computational tests showed both advantages and boundaries. When Gaussian noise was raised to a standard deviation of 30, the proposed framework’s AUC fell from 0.928 to 0.658, a relative decline of 29.1%. The comparative pose-based approaches declined by 34.8% to 38.2% under the same test, indicating greater relative robustness for the memory-enhanced system.

Removing the memory module produced a mean AUC of 0.831 with a 95% confidence interval of 0.820 to 0.842; configurations with 50, 100 and 200 memory units produced means of 0.875, 0.908 and 0.913, respectively, with the largest configuration offering only a marginal additional gain. The object detector contained 28.5 million parameters and required 45.2 gigaflops, while the behavior framework contained 35.8 million parameters. Its 54.1-fps inference rate corresponded to a reported total delay of 18.5 milliseconds per frame and a 4.5-GB GPU-memory footprint.

The improved detector was also tested beyond hospital footage. On VisDrone2021 and AI Challenger, which contain dense or small targets in aerial and street-view imagery, it achieved mAP@0.5 scores of 0.427 ± 0.011 and 0.502 ± 0.008, respectively. These values exceeded the original YOLOv8 by 4.5 and 5.1 percentage points and exceeded YOLOv9 by 2.0 and 2.6 points, while using 19.0% fewer parameters than YOLOv9 according to the study.

Sensitivity analysis also identified a smoke-detection peak at a detection-confidence threshold of 0.45, with 0.5 selected as a compromise across categories, and an optimal small-object weighting value of 0.7 in the tested range. A final test addressed the possibility that performance depended on one hospital’s cameras, angles and lighting. Models trained at Hospital A were evaluated directly, without fine-tuning, at a second institution, Hospital B, which differed in camera models, installation angles and illumination.

The detector’s mAP@0.5 declined from 0.918 to 0.872, a reduction of 4.6 percentage points, while cigarette-butt average precision was 0.781 and knife-holding accuracy was 0.854. The behavior framework’s mean AUC fell from 0.928 to 0.891, with AUCs of 0.886 for falls and 0.873 for smoking. The self-built data are available only on reasonable request, and no code is reported as available.

The authors identify additional scenario data, lightweight edge-device optimization and multi-view information as future directions for extending the approach.

Subject of Research: Computer-vision detection of hospital hazards and abnormal personnel behavior

Article Title: Hospital abnormal target and personnel behavior detection using improved YOLOv8

Article References: Li, Y., Yu, D., & Hu, B. (2026). Hospital abnormal target and personnel behavior detection using improved YOLOv8. Discover Artificial Intelligence, 6(1), Article 1052. https://doi.org/10.1007/s44163-026-02102-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02102-5

Keywords: hospital surveillance, YOLOv8, computer vision, small-object detection, video anomaly detection, pose estimation, optical flow, fall detection, smoking detection, knife detection, Hospital, abnormal