Vi-SAFE: Spatial‐Temporal Violence Detection

• The paper introduces Vi-SAFE, a spatial–temporal framework that decouples human localization via GE‑YOLOv8 from temporal violence recognition using TSN, achieving 0.88 ACC on RWF‑2000.
• It integrates GhostNetV3 with EMA attention and applies structured pruning, significantly reducing model complexity while maintaining detection precision in complex surveillance scenes.
• The modular design enables real-time deployment on edge devices by separating frame-level spatial detection from person-centric temporal analysis.

Vi-SAFE is a spatial–temporal violence detection framework for public surveillance that combines an optimized YOLOv8-style pedestrian detector, denoted GE‑YOLOv8, with a Temporal Segment Network (TSN). Its central design principle is to separate spatial localization from temporal reasoning—first localizing people frame by frame, then classifying violent versus non-violent behavior from person-centric temporal evidence—while maintaining a lightweight profile oriented toward real-time or near-real-time deployment on resource-constrained devices. The framework is evaluated on pedestrian detection data and on the RWF‑2000 surveillance benchmark, where it reports an accuracy of 0.88 and surpasses TSN alone at 0.77 (Chang et al., 16 Sep 2025).

1. Definition, scope, and operating assumptions

Vi-SAFE is intended for public surveillance scenarios such as stations, streets, and schools, where three constraints dominate: small-scale targets, cluttered and visually complex scenes, and the need for efficient temporal analysis over video streams. Its stated goals are high spatial precision for localizing humans involved in potential violent incidents, efficient modeling of motion and temporal dynamics over several seconds of video, and low computational cost in parameters and GFLOPs for edge deployment (Chang et al., 16 Sep 2025).

The framework is modular rather than jointly end-to-end. GE‑YOLOv8 is trained on a pedestrian detection dataset of 1,539 images, split 80/20 into training and testing sets, while TSN is trained on the RWF‑2000 violence dataset of 2,000 video clips, balanced as 1,000 violent and 1,000 non-violent, also with an 80/20 split. This separation means that Vi-SAFE operationalizes a pipeline in which human localization and violence recognition are optimized on distinct supervision sources before being connected at inference time (Chang et al., 16 Sep 2025).

At inference, for each frame $t$, GE‑YOLOv8 produces a set of pedestrian boxes

$$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$

and each box defines a cropped human region

$$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$

Over the video, TSN samples $K$ temporal segments from these person-centric regions and aggregates their features,

$$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$

to produce a video-level violent/non-violent decision. This design makes the temporal module explicitly person-centric rather than full-frame.

2. GE‑YOLOv8: spatial localization through lightweight detection

GE‑YOLOv8 starts from YOLOv8s and introduces three modifications: replacement of the standard backbone and head with GhostNetV3 components, insertion of an EMA attention mechanism in deeper backbone layers, and structured channel pruning guided by GroupNorm weights (Chang et al., 16 Sep 2025).

The GhostNetV3 substitution is motivated by the need to reduce parameters and FLOPs without materially degrading pedestrian localization. In the reported backbone comparison on the pedestrian dataset, vanilla YOLOv8s achieves Recall $0.709$, mAP $0.736$, $11.13$M parameters, and $28.4$ GFLOPs, whereas the GhostNetV3 replacement yields Recall $0.695$, mAP $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$0, $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$1M parameters, and $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$2 GFLOPs. The paper interprets this as roughly halving model complexity while leaving mAP nearly unchanged. MobileNetV4 and EfficientNetV2 are also reported, with GhostNetV3 described as offering the better accuracy–efficiency balance.

EMA attention is inserted into later backbone layers to improve feature extraction for small, occluded, or cluttered targets. The module uses horizontal and vertical average pooling, channel grouping, and spatial reweighting. Given a feature map $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$3, the module produces an attention map $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$4 and outputs

$$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$5

The practical effect described in the paper is that EMA highlights informative regions and suppresses background noise, improving detection of small or occluded targets and enhancing temporal consistency across frames.

A concise summary of the detector’s evolution is given below.

Configuration	Recall	mAP
YOLOv8s	0.709	0.736
YOLOv8s + EMA	0.696	0.741
YOLOv8s + GhostNetV3	0.695	0.732
YOLOv8s + GhostNetV3 + EMA	0.696	0.737

The attention ablation further shows GhostNetV3‑EMA at Recall $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$6 and mAP $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$7, exceeding GhostNetV3‑CBAM at $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$8 and GhostNetV3‑CoordAtt at $$B_t=\{(x_{t,i}^{1}, y_{t,i}^{1}, x_{t,i}^{2}, y_{t,i}^{2}) \mid i=1,\dots,N_t\},$$9. The reported compute remains $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$0M parameters and $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$1 GFLOPs for GhostNetV3‑EMA (Chang et al., 16 Sep 2025).

The final compression step is structured pruning. Channel importance is defined from GroupNorm weights: $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$2 Channels with the smallest $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$3 are pruned, after which the network is fine-tuned. The reported transition is from pre-pruning Recall $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$4, mAP $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$5, $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$6M parameters, and $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$7 GFLOPs to post-pruning Recall $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$8, mAP $$I_{t,i}=[y_{t,i}^{1}, y_{t,i}^{2}, x_{t,i}^{1}, x_{t,i}^{2}].$$9, $K$0M parameters, and $K$1 GFLOPs. The paper states that the accuracy drops are not statistically significant with $K$2.

3. TSN: temporal segment modeling for violence recognition

The temporal component is a Temporal Segment Network applied to cropped human regions rather than to the original full frame. TSN divides each video into $K$3 temporal segments, samples one frame or short snippet from each segment, applies a 2D CNN backbone to the sampled inputs, and aggregates segment-level evidence into a single video-level representation. The paper emphasizes that this strategy captures long-range temporal dependencies with substantially lower FLOPs than 3D convolutions (Chang et al., 16 Sep 2025).

Within Vi-SAFE, the sampled inputs are the ROIs produced by GE‑YOLOv8. If the sampled crops are denoted by $K$4, TSN outputs a global feature $K$5. The final classification layer is described as a linear classifier followed by a sigmoid, and the task is binary classification between violent and non-violent videos.

The RWF‑2000 dataset provides the temporal supervision. Because it consists of real-world surveillance footage with multiple scenes, lighting conditions, and crowd configurations, the temporal module is evaluated under conditions that already include heterogeneous backgrounds and interaction patterns. A key implication of the TSN design in this setting is that the temporal backbone need not model large amounts of irrelevant background motion once person-centric crops are supplied by the detector.

A common misconception is that Vi-SAFE is simply a stronger TSN baseline. The reported formulation is more specific: TSN alone is not the final system. The core contribution is the coupling of TSN with a person-localizing spatial front end, so that temporal reasoning is conditioned on detected human regions rather than on the surveillance frame as a whole.

4. Spatial–temporal coupling and inference workflow

The defining integration step in Vi-SAFE is the transfer of GE‑YOLOv8 detections into the TSN recognition pipeline. The workflow is: raw surveillance video is sampled into frames; GE‑YOLOv8 detects pedestrian boxes in each frame; each box is cropped into a person-centric ROI; the video is divided into $K$6 temporal segments; representative ROIs are sampled from those segments; TSN extracts and aggregates temporal features; and the final video-level violent/non-violent decision is produced (Chang et al., 16 Sep 2025).

This coupling has two immediate consequences. First, it reduces background interference by discarding large portions of the surveillance frame that are not associated with detected humans. Second, it makes the temporal classifier sensitive to human motion and interaction patterns rather than to scene-level appearance changes. The paper explicitly describes this as focusing TSN on humans and reducing background noise.

The pipeline is modular and asymmetrical. Spatial localization is frame-level and detector-driven; temporal reasoning is segment-level and classifier-driven. This suggests that improvements in either component can be propagated without redesigning the entire system. The paper itself notes that the modular design permits plugging in different detectors or temporal networks, though that is presented as a future direction rather than an evaluated variant.

The final output is a per-video decision, with potential per-frame ROIs available from the detector. The article does not claim a full spatio-temporal localization of the violent act itself; rather, it provides video-level classification assisted by human-region localization.

5. Experimental protocol, comparative performance, and efficiency

Experiments are reported under Ubuntu 18.04 with an AMD EPYC 9754 CPU, an NVIDIA RTX 4090D with 24 GB VRAM, Python 3.8, PyTorch 2.2.1, and CUDA 11.1. All results are averaged over 5 runs, and the paper reports ANOVA and Wilcoxon signed-rank testing for significance analysis (Chang et al., 16 Sep 2025).

For violence recognition on RWF‑2000, the comparative results are as follows.

Method	ACC
TSN	0.770
C3D	0.828
U‑Net + LSTM	0.820
ConvLSTM	0.770
TL	0.850
Openpose + ST‑GCN	0.878
Veltmeijer et al.	0.872
Vi-SAFE	0.880

The paper states that Vi-SAFE achieves $K$7 ACC and that the improvement is significant at $K$8 under one-way ANOVA. Relative to TSN alone, the improvement is $K$9 absolute. It also slightly exceeds Openpose + ST‑GCN, $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$0 versus $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$1.

On the detector side, the ablations isolate three contributions: GhostNetV3 as the lightweight backbone, EMA as the most effective attention mechanism among those tested, and structured pruning as the principal compression step. Together they define GE‑YOLOv8, which reaches mAP $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$2 pre-pruning and $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$3 post-pruning, with complexity reduced to $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$4M parameters and $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$5 GFLOPs after pruning.

The paper does not report explicit FPS or end-to-end latency. This suggests caution in interpreting “real-time” as a measured runtime claim. A more defensible reading is that the detector is compressed to $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$6M parameters and $$\mathcal{F}=\{f_1,f_2,\dots,f_K\}, \qquad \mathcal{G}=\mathrm{TSN}(\mathcal{F}),$$7 GFLOPs post-pruning, and the temporal module uses segment-sampled 2D CNN processing rather than 3D convolutions or RNN-heavy formulations. On that basis, the authors argue that the framework is suitable for resource-constrained edge deployment.

6. Limitations, future directions, and nomenclature

The stated limitations are primarily about generalization and difficult surveillance conditions. RWF‑2000, although real-world, still spans only a finite set of scenarios; the pedestrian dataset has 1,539 images and may not cover all camera locations, crowd configurations, and lighting conditions; extremely dense crowds and severe occlusions remain challenging; subtle violent behaviors with small motion signatures may be hard to distinguish from normal behavior; and deployment on very low-end hardware may still require additional optimization such as quantization or hardware-specific tuning (Chang et al., 16 Sep 2025).

The paper’s forward-looking statement is concise: future work will focus on improving generalization in complex environments by evaluating more diverse datasets and exploring advanced optimization strategies. A plausible implication is that the current system should be read as a modular baseline for efficient surveillance violence detection rather than as a closed final architecture.

The name also requires disambiguation. Vi-SAFE in this sense denotes the surveillance violence-detection framework of "Vi-SAFE: A Spatial-Temporal Framework for Efficient Violence Detection in Public Surveillance" (Chang et al., 16 Sep 2025). It should be distinguished from unrelated safety systems that use similar naming, including the UAV validation framework "ViVa-SAFELAND" (Soriano-García et al., 18 Mar 2025), the vision-only airborne collision-avoidance system "ViSafe" (Geng et al., 26 May 2025), the visual-inertial teleoperation safety filter for MAVs (Zhou et al., 2024), and the multi-vehicle reach-avoid planning formulation associated with double-obstacle Hamilton–Jacobi–Isaacs variational inequalities (Chen et al., 2014). The shared naming reflects a broader tendency to associate “Vi” with vision or viability and “SAFE” with safety enforcement, but these works address different technical problems, operating domains, and mathematical formalisms.