The team, led by Maya Houacine of the Laboratory of Research in Artificial Intelligence at the University of Science and Technology Houari Boumediene in Algiers, together with Khadidja Belattar and Naila Aziza Houacine, set out to correct what they see as persistent blind spots in the violence-detection literature. Many existing approaches lean on computationally heavy models that cannot plausibly run on the hardware guarding a real camera network. Others sidestep the timing question entirely, reporting accuracy without demonstrating that their systems can process video as it arrives. And a surprising number evaluate their methods on staged footage or cinematic fight scenes rather than the grainy, awkwardly angled, often distant video that real cameras actually produce. The Algerian group built its pipeline around real surveillance recordings and made real-time throughput a first-class design requirement rather than an afterthought.

The first stage of the system addresses a deceptively simple problem: video contains enormous redundancy. Consecutive frames in a surveillance feed are nearly identical, and training a neural network on every single frame wastes computation without adding information. The researchers compared two keyframe extraction techniques, methods that select the most informative frames from a clip while discarding duplicates, to determine which one best stripped redundancy while preserving the visual evidence of violent motion. Keyframe selection has a long history in video indexing, stretching back to early interfaces for browsing video archives in the 1990s, and modern variants rely on cues such as color histograms and adaptive clustering. Choosing the right extraction strategy matters because the quality of the frames fed into a detector directly shapes what the network can learn about fast, chaotic events like fights.

With preprocessing settled, the team adapted and evaluated three distinct deep learning architectures for the detection task. The first was an enhanced version of MobileNet, a convolutional neural network family originally engineered by Google researchers for mobile and embedded vision applications, where every parameter and every multiplication must be justified. The second combined MobileNetV2, which introduces inverted residual blocks and linear bottlenecks to squeeze more accuracy from fewer computations, with a long short-term memory network, a recurrent architecture designed to capture how visual features evolve across time, which is essential because violence is fundamentally a temporal phenomenon rather than a single frozen pose. The third contender was a three-dimensional convolutional neural network, or 3D CNN, which applies convolutions across both space and time simultaneously and is a popular choice for action recognition, though typically at a steep computational price.

Evaluation took place on two of the most challenging public benchmarks in the field. The first, RWF-2000, is a large-scale video database built specifically for violence detection using footage from real surveillance cameras, and it is openly available through the Zenodo repository. The second, UCF-Crime, comes from the University of Central Florida’s Center for Research in Computer Vision and contains real-world anomaly detection videos captured in uncontrolled settings, a subset of which the team used. The researchers supplemented these with material drawn from the Open Video Project archive. Testing on such data is deliberately unforgiving: lighting shifts, camera shake, occlusion, and the sheer variety of human behavior in public spaces all conspire to confuse models that were tuned on cleaner inputs.

A rigorous validation scheme separated this work from many predecessors. Rather than splitting the data once and reporting a single number, the team applied k-fold cross-validation to all three architectures, rotating through multiple partitions of the data so that performance estimates reflect the model’s behavior across different subsets rather than one lucky or unlucky split. This process exposed an unexpected complication in the enhanced MobileNet model: the folds were imbalanced, meaning some partitions contained uneven distributions of the classes the network needed to distinguish, which distorts both training and evaluation. Imbalanced folds are a subtle but consequential problem, because a model can appear to perform well simply because the data it sees is skewed.

The solution the researchers devised is one of the more inventive elements of the study. They turned to a genetic algorithm, a metaheuristic optimization method inspired by biological evolution, in which candidate solutions compete, mutate, and recombine over successive generations until a strong solution emerges. Their initial genetic algorithm reduced the fold imbalance cost, measured in misallocated images, from nearly 100 images down to 15. They then refined the approach into a dynamic genetic algorithm that adapted its behavior during the search, cutting the imbalance cost further to just 4 images while reaching the improved result in less time. The improvement demonstrates how classical optimization techniques can still play a decisive supporting role in modern deep learning pipelines, quietly fixing data plumbing problems that the headline models cannot solve on their own.

When the numbers settled, the MobileNet-based architecture emerged as the clear winner. It reached 90 percent accuracy, 91 percent precision, and 90 percent recall on the benchmark tasks, meaning it correctly identified violent incidents most of the time while keeping both false alarms and missed detections in check. Precision and recall matter in different ways for security applications: a system with poor precision floods operators with false alerts and quickly earns their distrust, while a system with poor recall lets genuine incidents slip through unnoticed. The MobileNet model balanced both while sustaining a processing speed of up to 130 frames per second, comfortably above the 25 to 30 frames per second that live video typically demands, leaving headroom for the other computational tasks a surveillance platform must perform.

The speed figure is the detail most likely to resonate with practitioners, because it addresses the gap between laboratory results and deployable systems. A detector that processes video slower than the cameras produce it accumulates an ever-growing backlog, and its alerts arrive after the moment to respond has passed. By building the system around efficient architectures and disciplined preprocessing, the Algerian team showed that state-of-the-art detection does not require data-center hardware. The choice of MobileNet as the backbone reflects a broader trend in computer vision toward models designed for edge devices, cameras, and embedded processors, where power budgets and memory constraints rule out the giant networks that dominate headline benchmarks.

The researchers were explicit about the ethical dimensions of their work, noting that surveillance data raises genuine privacy concerns. They mitigated these risks by relying exclusively on publicly available datasets, ensuring that no attempt was made to identify individuals, and confirming that no personally identifiable information was accessed or processed. All data were used solely for research purposes and handled in accordance with relevant privacy and data protection regulations, in line with principles of responsible data usage, data minimization, and respect for individual privacy. The code itself is not publicly available, but the datasets underpinning the study remain open, allowing other groups to reproduce the evaluation. As cities continue to multiply their camera networks and human attention stays fixed in supply, systems of this kind, fast enough to watch in real time and accurate enough to be trusted, are likely to move from research papers into the operational core of public safety infrastructure.

Subject of Research: Real-time violence detection in surveillance videos using lightweight deep learning models

Article Title: An intelligent system for real-time violence detection in surveillance videos

Article References: Houacine, M., Belattar, K., & Houacine, N. A. (2026). An intelligent system for real-time violence detection in surveillance videos. Multimedia Tools and Applications, 85(11), Article 815. https://doi.org/10.1007/s11042-026-21958-4

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21958-4

Keywords: violence detection, surveillance video, deep learning, MobileNet, LSTM, 3D CNN, keyframe extraction, genetic algorithm, cross-validation, computer vision, real-time systems, RWF-2000

The core insight of the new study is that the bottleneck in video search is no longer the accuracy of individual AI models but the lack of a coherent system that stitches them together. Deep multimodal models such as CLIP and its successors have made it possible to align images and text in a shared mathematical space, so that a sentence and a photograph of the same scene end up close together as high-dimensional vectors. Yet, as the authors argue, most research focuses on squeezing extra accuracy out of a single model while ignoring the messy engineering questions: how to chunk hours of raw footage, how to avoid indexing the same person thousands of times, how to keep track of who appeared when and where, and how to serve queries in milliseconds. Their answer is a two-part pipeline that cleanly separates offline indexing from online retrieval.

On the indexing side, the system ingests raw video files and first splits them into roughly ten-minute chunks using FFmpeg, preserving the original encoding so each segment can be processed independently and in parallel. A frame-sampling module then skips redundant frames, since typical surveillance footage recorded at 30 or 60 frames per second offers far more temporal detail than person search actually needs. The sampled frames flow into YOLO11, the latest generation of Ultralytics’ real-time object detector, chosen for its favorable speed-accuracy trade-off, paired with the BoT-SORT tracker, which combines motion cues with appearance features to assign each detected person a persistent identity across frames. On the MOT17 benchmark, BoT-SORT reports 80.5 MOTA and 80.2 IDF1, meaning it rarely confuses one pedestrian with another, a property the authors consider essential for generating temporally coherent metadata.

Detection alone would still flood the database with near-duplicate images, so the pipeline adds a quality-filtering stage that is among the paper’s most practical contributions. Every cropped person image receives a score combining the detector’s confidence with an edge penalty: bounding boxes that touch the frame border, indicating a partially visible subject, are down-weighted. A perceptual hashing step then compares consecutive crops within the same tracked trajectory, discarding any whose hash differs by fewer than a threshold number of bits, a signal that the image is essentially identical to the previous one. What survives is a compact, diverse gallery of high-quality person crops rather than a bloated collection of blurry, truncated, or repetitive snapshots.

Each surviving crop is then transformed into a semantic fingerprint by SigLIP2, a multilingual vision-language encoder that extends the original SigLIP architecture with a sigmoid-based training loss, self-supervised objectives, and careful data curation. Because the model was trained on paired images and text, its visual embeddings live in the same latent space as its text embeddings, which is precisely what allows a typed sentence to retrieve a photograph. The system also computes a second, complementary embedding for each person using InsightFace: the RetinaFace detector localizes the face and its landmarks even under challenging poses and lighting, and the ArcFace recognizer converts the aligned face patch into a highly discriminative identity vector. Both embeddings, along with rich metadata such as source video, timestamps, track identity, and bounding-box coordinates, are stored in a Qdrant vector database, so every search result can be traced back to the exact frame in the original footage.

At query time, the retrieval pipeline mirrors the indexing representation. A user can submit an image, a natural-language description, or both; the same SigLIP2 encoders embed the inputs, guaranteeing that queries and database entries are directly comparable. When both modalities are present, a tunable balance parameter blends the image and text vectors, letting users slide between purely visual similarity and purely linguistic matching. A face-based mode restricts the search to identity embeddings when a face is detectable, and a negative-prompt field excludes unwanted attributes. The back-end runs an approximate nearest-neighbor search with cosine similarity, then consolidates results so that only the single best match per tracked object per video is returned, each accompanied by the retrieved frame, a similarity score, and metadata including camera identifier, first-appearance timestamp, and duration in the scene.

The quantitative centerpiece of the paper is an evaluation on RSTPReid, a benchmark of 20,505 pedestrian images covering 4,101 individuals captured by 15 cameras, each image paired with two text descriptions. The researchers fine-tuned the SigLIP2-SO400M-Patch14-384 checkpoint on a curated mixture of person-centric image-text datasets, including CUHK-PEDES, ICFG-PEDES, RSTPReid itself, an attribute-rich multiview re-identification set, and the video-based TVPReid corpus, totaling roughly 316,000 training pairs. Training used the native contrastive loss on four NVIDIA A100 GPUs on the Cineca Leonardo supercomputer, with early stopping based on validation mean Average Precision. The result: a mean Average Precision of 0.68, comfortably ahead of the best competing value of 0.54 among recent text-based person search methods such as OCDL, UP-Person, RDE, CTGI, WoRA, MARS, MRA, and CONQUER, while Recall@1 remained competitive within one standard deviation of the leaders.

Notably, the gains did not come from inventing new architectures. When the model was fine-tuned only on the three standard datasets used by the comparison methods, it already achieved the best mAP, indicating that the SigLIP2 backbone itself provides a stronger shared embedding space for matching descriptions to pedestrian attributes. Adding the larger, more diverse corpus then improved all metrics further. Ablation experiments on a subset of the Wildtrack multi-camera surveillance dataset underscored the value of the system’s design choices: removing face embeddings dropped Recall@1 from about 51 percent to 33 percent, while replacing quality-scored crops with uniformly sampled ones collapsed mAP from 0.66 to 0.39, showing that both identity cues and careful crop selection matter substantially.

The scalability analysis offers a sober, engineering-minded picture. Processing 128 ten-minute videos, about 256 gigabytes of footage, took roughly 90 minutes on a four-GPU server, with tracking saturating the CPU and encoding bottlenecked by memory traffic and model replication across worker processes rather than raw computation. Extrapolating linearly, the authors estimate the current setup could index about a terabyte of video in six hours, though they are candid that GPU utilization during encoding remains low and that a centralized inference server would be needed for truly efficient large deployments. They also acknowledge limitations: the system is person-centric, does not yet model complex temporal activities or events, and the auxiliary training data included AI-generated descriptions that were only spot-checked rather than fully validated by humans.

Even with those caveats, the study reads as a template for how modern AI systems should be built: not as isolated models chasing leaderboard numbers, but as orchestrated pipelines where detection, tracking, filtering, embedding, metadata, and search reinforce one another. The authors envision extending the framework beyond people to vehicles and other objects, adding temporal and multi-view reasoning, and eventually integrating large language models and knowledge graphs for conversational, explainable video search. If surveillance archives, streaming platforms, and personal devices keep growing at their current pace, systems that turn unwatchable oceans of footage into a simple search box may soon become as unremarkable, and as indispensable, as the web search bar itself.

Subject of Research: A scalable semantic pipeline for video indexing and text-based person retrieval using multimodal vision-language embeddings

Article Title: A scalable and semantic pipeline for efficient video indexing and person retrieval

Article References: Hmaidan, R., Milardo, S., Donato, I., Greco, D., Ingargiola, A., & Reforgiato Recupero, D. (2026). A scalable and semantic pipeline for efficient video indexing and person retrieval. Multimedia Tools and Applications, 85(9), Article 727. https://doi.org/10.1007/s11042-026-21882-7

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21882-7

Keywords: video retrieval, person re-identification, multimodal embeddings, SigLIP2, YOLO11, BoT-SORT, face recognition, vector database, surveillance video, computer vision, text-based person search, RSTPReid

The approach, called RMTA-Net, belongs to a family of techniques known as prediction-based unsupervised video anomaly detection. The core idea is elegant. The network is shown a short clip of video frames and asked to predict what the next frame will look like. Because it is trained only on normal scenes, it becomes an expert at forecasting ordinary motion and appearance. When an anomalous event unfolds, the model’s prediction diverges sharply from the actual frame, and the size of that discrepancy—measured pixel by pixel—becomes the anomaly score. Frames that deviate strongly from expectation are flagged for human review. This sidesteps the need for exhaustive annotation, since the model learns the statistics of normality from unlabeled footage, but it places enormous demands on the network’s ability to represent both what objects look like and how they move through time.

Existing prediction-based methods, the authors argue, fall short on two fronts. First, their spatial representations—the internal descriptions of shapes, textures, and objects in each frame—are often too coarse to capture subtle appearance variations. Second, they struggle to understand clip-level temporal dependency: the way motion in one frame conditions and constrains motion in the next. Complex scenes contain overlapping movements, occlusions, and camera perspective shifts that make naive temporal modeling brittle. A pedestrian partially hidden behind a van, a crowd flowing in two directions at once, or a vehicle entering the frame at high speed all require the model to reason jointly across space and time rather than treating each frame in isolation. RMTA-Net was designed specifically to close both gaps within a single architecture.

The first stage of the pipeline is a residual spatial feature enhancement network. Operating on individual frames, it progressively extracts structural and appearance information at multiple levels of abstraction, from fine edges and textures to coarse object layouts. The residual design means that each processing stage refines the features produced by the previous one, adding detail rather than replacing it, which helps preserve the subtle cues that distinguish a person from a shadow or a stroller from a bicycle. These independently encoded spatial features are then organized into an explicit temporal representation, effectively stacking the per-frame descriptions into a sequence that downstream modules can interrogate for motion patterns and inter-frame relationships.

The heart of the system is the module that gives the network its name: recurrent conditioned memory-guided temporal attention, or RMTA. This component fuses two ideas that have usually been pursued separately. The first is recurrent temporal processing, in which a recurrent network maintains a running internal state as it steps through the frames of a clip, carrying forward information about what has already happened. The second is a set of learnable memory banks—trainable storage that encodes dataset-level normality priors, essentially compressed prototypes of how normal scenes tend to look and move. During processing, the model performs attention-based memory access, querying the memory banks with the recurrent network’s output to retrieve relevant normality prototypes, while a parallel branch uses temporal attention to retrieve a global normality prior driven by the contextual relationships among frames in the clip.

Crucially, the two branches do not simply compete; they are combined through adaptive gating, a learned mechanism that decides, at each moment, how much weight to give the recurrent memory retrieval versus the contextual temporal relationships. When motion is regular and predictable, the recurrent pathway may dominate; when the scene is complex or the recurrent state is uncertain, the memory-guided global prior can compensate. This dual-branch design allows the model to maintain a robust sense of normality even under the appearance variations and tangled motion dynamics that degrade earlier methods. The fused representation is then passed to an attention-enhanced decoder, which reconstructs the future frame from the learned spatiotemporal embeddings. The decoder’s attention mechanism helps it focus on the regions of the scene most relevant to the prediction, sharpening the contrast between well-forecast normal content and the surprising content that signals an anomaly.

The authors evaluated RMTA-Net on three widely used benchmark datasets that together span the practical challenges of surveillance video. UCSD Ped2 contains pedestrian walkways filmed from a fixed viewpoint, with anomalies such as cyclists and small vehicles intruding into pedestrian-only spaces. CUHK Avenue features a campus avenue with more varied activities, including loitering, throwing objects, and unusual walking patterns. ShanghaiTech is the most demanding of the three, with thirteen scenes of differing complexity, crowded conditions, and diverse anomaly types. On these benchmarks, RMTA-Net achieved frame-level area-under-the-curve scores of 99.0 percent on UCSD Ped2, 90.1 percent on CUHK Avenue, and 75.8 percent on ShanghaiTech, remaining competitive with several state-of-the-art methods across the board. The AUC metric reflects how well the system ranks anomalous frames above normal ones, so scores approaching 99 percent on Ped2 indicate near-perfect discrimination in that relatively controlled setting, while the ShanghaiTech result demonstrates useful performance in genuinely cluttered, multi-scene environments.

What makes this result notable is not merely the headline numbers but the architectural lesson embedded in them. Memory-based reasoning gives the network something like an institutional memory of normality: rather than relying solely on the hidden state of a recurrent network, which must compress an entire clip into a fixed-size vector, the model can consult explicit prototypes of normal appearance and motion. Attention-guided retrieval means those prototypes are consulted selectively, only where the current context calls for them. The adaptive gate then arbitrates between memory and context on the fly. This division of labor mirrors, in a loose computational sense, the way long-term and episodic memory support expectations in human perception, a parallel the authors’ choice of terminology makes explicit. The result is a system whose predictions degrade gracefully under complexity rather than collapsing when scenes become crowded or visually ambiguous.

The practical implications extend well beyond benchmark scores. Because the method requires no anomaly labels, it can be deployed on the vast reservoirs of unlabeled footage that organizations already possess, learning the normal rhythm of a specific environment and flagging deviations without prior knowledge of what those deviations might be. That matters for applications ranging from public safety and traffic monitoring to industrial safety and elder care, where the set of possible incidents cannot be enumerated in advance. The open-access publication, with fees covered by the Government of India’s One Nation One Subscription initiative, means the full technical description is freely available to researchers and practitioners worldwide, lowering the barrier for replication and extension.

Challenges remain, as the authors’ own results make clear. Performance on ShanghaiTech, while competitive, still leaves substantial room for improvement in the hardest scenes, and future-frame prediction systems in general can be sensitive to camera jitter, illumination shifts, and the sheer diversity of real-world behavior. The field will also need to keep examining how such models behave across different populations and environments to ensure that flagged anomalies reflect genuine events rather than artifacts of the training distribution. Nevertheless, RMTA-Net offers a compelling demonstration that combining recurrent temporal modeling, learnable memory, and adaptive attention yields a more faithful model of normality—and that the better a machine understands the ordinary, the more sharply the extraordinary stands out. In surveillance, where the abnormal is rare by definition, that principle may prove to be the key to systems that finally earn the trust placed in the cameras they watch.

Subject of Research: Unsupervised video anomaly detection using future-frame prediction with memory-guided temporal attention

Article Title: RMTA-Net: recurrent conditioned memory-guided temporal attention with spatial feature enhancement for video anomaly detection

Article References: Chouhan, S. P., Punn, N. S., & Bhattacharya, M. (2026). RMTA-Net: recurrent conditioned memory-guided temporal attention with spatial feature enhancement for video anomaly detection. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02506-x

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02506-x

Keywords: video anomaly detection, future-frame prediction, recurrent neural networks, memory banks, temporal attention, spatial feature enhancement, unsupervised learning, surveillance video, computer vision, deep learning, UCSD Ped2, ShanghaiTech