Meta-Learning Approach for Adaptive Anomaly Detection from Multi-Scenario Video Surveillance

Abstract

Video surveillance is widely used in different areas like roads, malls, education, industries, retail, parks, bus stands, and restaurants, each presenting distinct anomaly patterns that demand specialized detection strategies. Adapting anomaly detection models to new camera viewpoints or environmental variations within the same scenario remains a significant challenge. Extending these models to entirely different surveillance environments or scenarios often requires extensive retraining, which can be both resource-intensive and time-consuming. To overcome these limitations, model frameworks, i.e., the video anomaly detector model, have been proposed, leveraging the meta-learning framework for faster adaptation using swin transformer for feature extraction to new concepts. In response, the dataset named MSAD (multi-scenario anomaly detection) having 14 different scenarios from multiple camera views, is the high resolution anomaly detection dataset that includes diverse motion patterns and challenging variations such as varying lighting and weather conditions, offering a robust foundation for training advanced anomaly detection models. Experiments validate the effectiveness of the proposed framework, which integrates model-agnostic meta-learning (MAML) with a ten-shot, one-query adaptation strategy. Leveraging the swin transformer as a spatial feature extractor, the model captures rich hierarchical representations from surveillance videos. This combination enables rapid generalization to novel viewpoints within the same scenario and maintains competitive performance when deployed in entirely new environments. These results highlight the strength of MAML in few-shot learning settings and demonstrate its potential for scalable anomaly detection across diverse surveillance scenarios.

1. Introduction

Anomaly detection is essential in fields such as surveillance, healthcare, finance, and industrial monitoring, where identifying deviations from normal patterns is crucial. In video surveillance, anomalies can include unusual movements, unauthorized access, or unexpected incidents, with context playing a significant role in determining what is considered abnormal, for example, walking on a highway instead of a sidewalk. Since obtaining labeled video data for anomalies is challenging, many existing methods approach anomaly detection in video surveillance as a one-class classification problem, where models are trained solely on normal activity and expected to flag deviations as anomalies during testing. However, this approach presents several critical challenges when applied to complex, real-world environments. First, existing models often lack robustness to environmental variability, struggling to handle conditions such as changes in lighting, weather, or camera motion—factors commonly encountered across surveillance scenarios like highways, parks, offices, and warehouses. Second, the limited diversity of current datasets can result in poor anomaly representation, causing common activities like running or cycling on a road or sidewalk to be misclassified. Third, most anomaly detection frameworks are human-centric, focusing mainly on activities like assault, fighting, or robbery, while overlooking non-human anomalies such as fires, falling objects, explosions, water leaks, or traffic accidents. Fourth, a significant gap exists in scenario diversity: many datasets include only different camera angles within a single environment, lacking the cross-scenario variability needed to train generalizable models.

To address these gaps, an MSAD dataset [1] was used that incorporates 11 diverse anomaly types—including assault, explosion, fighting, fire, objects falling, people falling, robbery, shooting, traffic accident, vandalism, and water incidents—across 14 distinct normal surveillance scenarios such as front door areas, highways, malls, offices, parks, restaurants, parking lots, pedestrian streets, roads, shops, sidewalks, street highways, trains, and warehouses. This combination introduces substantial variability in both scene context and anomaly nature, underscoring the need for adaptive approaches like few-shot meta-learning that can effectively generalize across such a wide range of spatial patterns.

In addition to the dataset, this study uses MAML [2,3] as a novel approach to anomaly detection. Meta-learning, or “learning to learn,” enables models to quickly adapt to new tasks with minimal data, which is particularly valuable when large-scale data collection is impractical. The meta-learning process consists of two phases: the meta-training phase, where the model is exposed to various tasks (i.e., different scenarios in the MSAD dataset) to develop a meta-model capable of adapting to new tasks, and the meta-testing phase, where the model is evaluated on previously unseen tasks to assess its ability to generalize.

This meta-learning approach offers several advantages for anomaly detection. It allows models to rapidly adapt to new environments with minimal data, improving their ability to function in dynamic settings. Training in diverse tasks enhances the model’s ability to generalize, making it more effective in identifying anomalies in unfamiliar scenarios. Additionally, meta-learning increases efficiency by leveraging small amounts of anomalous data, addressing the common issue of data scarcity. Applying MAML to the MSAD dataset enables the model to handle new viewpoints by learning from multiple camera perspectives within the same scenario. It also demonstrates strong generalization capabilities when applied to entirely new scenarios, proving its effectiveness in real-world applications. This approach significantly strengthens anomaly detection in dynamic and evolving environments, making it a promising solution for improving security and monitoring systems.

Previous studies on anomaly detection across multiple scenarios have faced significant challenges in generalization. Traditional models struggle to adapt to different scenarios and viewpoints, and while few-shot learning (FSL) [1] methods have been introduced to address this issue, they remain limited to specific environments and fail to generalize effectively. To overcome these limitations, a meta-learning framework, particularly MAML, can be employed. MAML optimizes model parameters to enable rapid adaptation to new tasks with only a few examples through minimal gradient updates.

The main contributions of this paper can be summarized as follows:

• The MAML technique is designed to enable anomaly detection in 14 different scenarios, including highways, shops, front doors, offices, parking lots, pedestrian streets, street highviews, warehouses, roads, parks, malls, trains, restaurants, and sidewalks by using a small number of normal and anomalous data, where data collection is limited.
• Our method leverages metadata and swin transformers to extract spatial features from frames of video data from different targeted scenes, thereby enhancing adaptability to different scenarios.
• By leveraging anomaly detector models using MAML to acquire the capability to classify between normal and abnormal data, better accuracy in anomaly detection from different scenarios with limited data is realized from the targeted domain.

2. Related Works

Over the years, several counter measures have been proposed to detect anomalies on video surveillance in different sectors like education, manufacturing, roads, etc. There are two different sub-sections that include deep learning and anomaly detection works.

2.1. General Deep Learning Methods

Over the years, many supervised deep learning methods, including convolutional neural networks (CNNs), recurrent networks, and transformers, have shown strong performance in computer vision tasks. However, their effectiveness depends on access to large, labeled datasets, which are often costly and time-consuming to create, especially in multi-scenario surveillance systems with limited anomalous data.

Hwang et al. 2025 [4] introduced the work that exemplifies how augmentation strategies can enhance performance when labeled data is scarce. While such techniques help, they cannot fully eliminate the dependency on annotated datasets, underscoring the need for more adaptive learning frameworks.

Park et al. 2025 [5] investigated the effectiveness of multiple image augmentation techniques—such as brightness, contrast, and rotation—on non-PPE detection in construction site images using the YOLOv8 model. Their findings showed that well-chosen augmentations can significantly enhance detection accuracy, especially for classes with limited data. Their study focused on static image classification.

Santoro et al. 2016 [6] focused on meta-learning techniques that address this by enabling models to learn from prior tasks and adapt quickly to new ones. Memory-Augmented Neural Networks (MANNs), especially Neural Turing Machines, provided external memory mechanisms that support rapid learning from a few examples. They also demonstrated that MANNs outperform LSTMs in both classification and regression tasks.

Finn, Abbeel et al. 2017 [2] concluded that MAML is a powerful and versatile meta-learning approach that allows models to adapt quickly to new tasks with minimal data. Its model-agnostic nature makes it applicable across various domains and tasks, addressing key limitations of traditional machine learning methods in dynamic environments.

Ravi et al. 2017 [7] studied and introduced an LSTM-based meta-learner that learns an optimization strategy tailored for few-shot scenarios by efficiently updating a learner model’s parameters.

Munkhdalai et al. 2017 [8] introduced Meta Networks (MetaNet), a novel meta-learning framework designed to enhance rapid generalization and continual learning in neural networks, particularly in scenarios with limited labeled data.

Joshi et al. 2023 [9] focused on the head pose estimation techniques, in which they compared MAML and FO-MAML methods for efficiency. FO-MAML uses first-order approximation, avoiding Hessian computation. Experiments were conducted in one-shot, five-shot, and ten-shot settings. FO-MAML showed faster optimization than traditional MAML.

2.2. Anomaly Detection Methods

Gong et al. 2019 [10] studied reconstruction-based techniques, such as those employing Auto-Encoder methods exclusively trained on normal videos and subsequently tested on both normal and abnormal videos.

Bergman et al. 2020 [11] presented a novel approach to anomaly detection through a classification-based method called GOAD, which addressed the limitations of existing techniques by utilizing a semi-supervised learning framework. It was built on previous works that categorize anomaly detection methods into reconstruction, distributional, and classification-based approaches, highlighting the effectiveness of deep learning in this domain.

Chen et al. 2021 [12] studied anomaly detection as a one-class classification problem, relying on the features of normal events where a generative adversarial network aims to reconstruct normal events to minimize the reconstruction loss.

Duan et al. 2024 [13] introduced a meta-learning framework for efficient anomaly diagnosis in few-shot AIOps scenarios, addressing the challenges of limited labeled data in IT operations.

Jeong et al. 2024 [14] presented a novel approach for detecting pipeline leaks using limited normal data, particularly in hazardous environments like power plants. They proposed the GAD-PN framework, combining Cycle GAN for generating synthetic anomaly data and Prototypical Networks for anomaly classification based on few-shot learning.

Li et al. 2023 [15] proposed a novel zero-shot anomaly detection method leveraging batch normalization statistics to detect anomalies without requiring training data from the target distribution. Their approach models the distribution of batch normalization parameters across diverse tasks, enabling effective generalization to unseen domains.

Meng et al., 2023 [16] proposed an explainable few-shot learning framework for online anomaly detection in ultrasonic metal welding processes with varying configurations. Their approach combined prototype-based learning with interpretability techniques to enable accurate detection and transparent decision making using limited labeled data.

Moon et al. 2023 [17] proposed an anomaly detection method using a model-agnostic meta-learning-based variational autoencoder (MAVAE) tailored to facility management systems. This approach enables rapid adaptation to new environments using few-shot learning by leveraging MAML to fine-tune VAE parameters with limited normal data.

Navarro et al. 2023 [18] introduced Hydra, a meta-learning framework designed for fast model recommendation in unsupervised multivariate time series anomaly detection. Their approach leverages a small set of meta-features to match datasets with suitable detection models, addressing the challenges of scalability and generalization.

Wu et al. 2021 [19] introduced an unsupervised transformer-based framework, Meta Former, for anomaly detection without relying on labeled data. The model leverages self-attention mechanisms to learn generalized representations from normal patterns, enabling it to detect anomalies across diverse domains.

Reiss et al. 2022 [20] studied and experimented with anomaly detection methods, which were often thought to struggle with understanding normal events, although some models can effectively detect anomalies. However, achieving a good reconstruction is not a guarantee of effective anomaly detection. These methods can be sensitive to object speeds, leading them to misclassify sudden motions as anomalies.

Soh and Cho et al. 2020 [21] published a paper that introduced a meta-transfer learning framework that could be adapted for anomaly detection tasks. The method aims to learn a meta-learner that can transfer knowledge from related tasks to new tasks with limited labeled data.

Smith et al. 2020 [22] addressed the limitations of traditional anomaly detection methods in cybersecurity, emphasizing the potential of meta-learning to enhance detection accuracy and adaptability with limited data. Meta-learning outperformed traditional models, especially with limited data. Initial training was resource-intensive, but adaptation to new tasks was efficient.

Natha et al. 2022 [23] conducted a systematic review of anomaly detection methods utilizing machine learning and deep learning techniques across various domains. The study categorized approaches based on supervised, unsupervised, and semi-supervised learning, highlighting their strengths, limitations, and application areas.

Zhang et al. 2020 [24] proposed a novel few-shot learning framework leveraging MAML, which was proposed for bearing anomaly detection. The framework aimed to develop a robust fault classifier with minimal data by adapting quickly to new fault conditions. In a case study involving the introduction of new artificial faults, the method achieved up to 25 percent overall accuracy compared to a Siamese network benchmark. Moreover, when sufficient training data from artificial damages was available, MAML demonstrated competitive generalization abilities compared to state-of-the-art few-shot learning methods in identifying realistic bearing damages.

Zhu, Raj et al. 2024 [1] developed a SAAD (Scenario Adaptive Anomaly Detection) model that takes a unique approach, utilizing a few-shot learning framework for faster adaptation to novel concepts and scenarios. This adaptability, coupled with its competitive performance on diverse scenarios, sets SAAD apart from traditional models.

Wu et al. 2024 [25] provided a comprehensive review of deep learning techniques for video anomaly detection, highlighting advancements across supervised, unsupervised, and self-supervised methods. The paper categorized key approaches based on learning paradigms, feature representations, and anomaly scoring techniques.

Zhang et al. 2024 [26] proposed ADAGENT, a novel anomaly detection framework that leverages multimodal large models to operate effectively in adverse and complex environments. By integrating visual, textual, and contextual data, ADAGENT enhances anomaly detection robustness under challenging conditions such as low visibility or sensor noise.

Zhu, Wang et al. 2024 [27] proposed the Scenario-Adaptive Anomaly Detection (SA2D) model to enhance few-shot learning adaptability across diverse scenarios and viewpoints. Leveraging a meta-learning framework, SA2D effectively adapts to unseen contexts. It uses future frame prediction for anomaly detection, supported by U-Net and ConvLSTM for temporal modeling and a GAN for frame reconstruction. The authors also introduced the MSAD dataset, featuring frame-level annotations across varied indoor/outdoor scenes and multiple camera angles, addressing the limitations of datasets like UCSD Ped1/Ped2 and Shanghai Tech (Shanghai, China). The study emphasized the role of scenario diversity and meta-learning in advancing video anomaly detection.

In contrast to existing methods, meta-learning framework can be utilized in models to adapt to multiple scenarios and even new scenarios with more accuracy. This includes MAML, which trains a model’s parameters in such a way that a small number of gradient updates can quickly adapt to a new task with few examples. This approach is versatile and can be applied to various domains, including image classification, regression, and reinforcement learning.

3. Methodology

The methodology starts with data collection. The system model for video anomaly detection using MAML from video surveillance is illustrated in Figure 1, where frames were extracted and then preprocessed.

After preprocessing, feature extraction was done using swin transformer. The extracted features as feature maps were fed to the anomaly detector model, which used MAML to classify the dataset into normal and anomalous.

3.1. Dataset

MSAD [1] datasets containing anomalies and normal instances were collected. The MSAD dataset has anomaly videos, normal testing, and normal training datasets in video form. The dataset also contains annotated csv files that contain frame numbers. The CSV file dataset has a column containing the range of values indicating when anomalies began. The MSAD dataset, along with any new datasets, serves as the input to the system. Here, each dataset represents a different task.

A collection of 720 video datasets was compiled, categorized as follows: anomaly videos, normal testing videos, and normal training videos. The anomaly video dataset comprises 240 videos, encompassing 11 distinct anomaly types across 13 different scenarios, excluding highway scenes. The normal testing video set includes a total of 120 videos. Finally, the normal training video dataset contains 360 videos covering 14 different scenarios, including highway scenes.

3.2. Data Preprocessing

Raw data from the MSAD dataset underwent preprocessing to ensure data quality and consistency. This involved video frame resizing using a bilinear interpolation method and normalization using means and standard deviations. Here, bilinear interpolation [28] for y (height) and x (width) contains four coordinate points A(y₁, x₁), B(y₁, x₂), C(y₂, x₁), and D(y₂, x₂). A formula for computation of X and Y for width dimension is given by:

X = A(1 − w_x) + Bw_x

(1)

Y = C (1 − w_x) + Dw_x,

(2)

Then linear interpolation between the two interpolated values X and Y in the height dimension is given by:

Z = X (1 − w_y) + Yw_y,

(3)

= A(1 − w_x)(1 − w_y) + Bw_x(1 − w_y) + C (1 − w_x)w_y + Dw_xw_y

(4)

where wx = (x − x₁)/(x₂ − x₁) and w_y = (y − y₁)/(y₂ − y₁)

Using the above equations, images were resized to the required dimension, i.e., a 224 × 224 RGB image. Here, PIL images with [H × W × C] in range [0, 255] with an unsigned int type were converted to Tensor of float type with shape [C × H × W] in the range of [0, 1]. In this case, PIL images belong to RGB mode. The normalization of resized images was done to standardize the images based on their mean and standard deviation, ensuring stable and consistent model performance.

For this study, the dataset was initially partitioned into training, validation, and testing sets based on different scenarios. For the initial training phase of the model, nine distinct scenarios were utilized, encompassing shops, front doors, offices, parking lots, pedestrian streets, street highviews, warehouses, roads, and parks, each containing both anomalous and normal video data. Two different scenarios, malls and trains, which also included both anomalous and normal videos, were allocated for the initial validation of the model. Finally, the initial testing of the model employed data from two separate scenarios, restaurants and sidewalks, both containing anomalous and normal video examples.

3.3. Feature Extraction

For video frames (Image Input) of size 224 × 224 with 3 channels, the swin transformer [29] was used to extract feature vectors from the preprocessed data of the MSAD dataset, which is illustrated in Figure 2.

Here, swin transformers were used for leveraging their hierarchical attention mechanism to capture both local and global spatial patterns. During the forward pass, the model processed the input images, and feature maps were extracted from intermediate layers, providing rich representations for anomaly detection. For mathematical interpretation of input image I, the feature extraction process using Swin-S can be described as:

F = SwinS(I),

(5)

where I is the input image of size 224 × 224 × 3 and F is the feature vector whose output feature shape is 768 dimensional vectors (feature vectors). In patch embeddings, the image I is divided into patches of size 4 × 4, and each patch is embedded into a patch token T as

T = W_EI + b_E,

(6)

where W_E is the learnable embedding weight and b_E is the bias. Then, for hierarchical feature representation, the swin transformer applies Shifted Windows Multi-Head Self-Attention (SW-MSA) in hierarchical stages using:

F(l) = SW-MSA(MLP(F^(l−1))),

(7)

where F^(l) is the feature map at layer l and SW-MSA is the shifted window-based multi-head self-attention. Then, feature extraction at final layer was performed, where final feature representation F can be taken from the last stage of the swin transformer using:

F = Pooling(F^(L)),

(8)

where L is the last layer and the pooling operation reduces spatial dimensions to a feature vector.

Finally, the input image 224 × 224 × 3 was given to the swin transformer, and 768 dimensional vectors as feature vectors were obtained. These embeddings then served as inputs to the anomaly detector model, allowing for effective anomaly detection.

3.4. Anomaly Detector Model

In this phase, the architecture represents a feedforward neural network designed for anomaly detection using image embeddings as the input. The model begins with an input layer that processes these embeddings, which are typically extracted from the swin model. Figure 3 shows the layered architecture, where input image embeddings (dimensional vectors) are given to the input layer. The anomaly detector class takes image embeddings as the input, with a default feature size of 1024. The model consists of three sequential layers, each containing a fully connected (Linear) layer, batch normalization (BatchNorm1d), and a ReLU activation function to introduce non-linearity. The first layer reduces the input dimension from 1024 to 512, followed by a second layer that maintains the dimension at 512, and a third layer that reduces it further to 256. The final output layer is a linear transformation that maps the 256-dimensional feature representation to a single scalar logit. The forward method processes an input tensor through these layers sequentially, producing an output shape, where each sample gets a single logit value. This output is used for binary classification, where we can get a probability for determining whether an input is normal or anomalous.

3.4.1. Meta Training on Dataset

The meta training process in meta-learning consists of several key steps designed to help a model learn how to quickly adapt to new tasks (image embeddings) with limited data. A meta-learning approach is illustrated in Figure 4.

• Task Sampling: The sample task method was designed to generate a dataset sample for an anomaly detection task. Here, a random scenario from a predefined list was selected. Within that scenario, it identified normal and anomalous images by cross-referencing scenario-specific indices with label-specific indices. Once these images were identified, it randomly selected a fixed number (k_shot + k_query) of normal and anomalous samples. These selected images were then split into two sets: the support set, which consisted of k_shot normal and k_shot anomalous images for training, and the query set, which contained k_query normal and k_query anoma-lous images for evaluation. Each image was assigned to its respective label, either normal or anomalous. To avoid bias, the method shuffled both the support and query sets before returning them. Then, the function ultimately ensured a balanced, struc-tured dataset that can be used in meta-learning tasks, particularly for detecting anomalies within specific scenarios.
• Inner Loop: The method began by creating a temporary copy of the model to ensure that updates do not affect the original model. The copied model was then placed in training mode, and a Binary Cross Entropy with Logits Loss function was defined for classification. Then, a series of inner-loop updates was performed over self-inner steps iterations. In each iteration, it made predictions using the support set, calculat-ed the loss, and computed gradients with respect to the model’s parameters. Instead of using an optimizer, it manually updated each parameter using stochastic gradient descent (SGD) by subtracting the gradient scaled by the inner learning rate. This adapted model, then fine-tuned to the support set, was returned. The purpose of this method was to allow a model to quickly adapt to a new task using only a few exam-ples. This approach maintained the computational graph so that the outer update can backpropagate through the inner adaptation steps.
• Outer Loop: The meta update function implements a meta-learning step based on the MAML approach. It processed multiple tasks, where each task consisted of a support set (used for inner-loop adaptation) and a query set (used for evaluation). First, it ini-tialized the loss function, optimizer, and storage variables for query labels and pre-dictions. Then, it iterated over tasks, moving data to the appropriate device and per-forming an inner-loop update, where the model adapted to the task using the support set. After adaptation, the model made predictions on the query set, and the binary cross-entropy loss was computed. The predicted probabilities and true query labels were stored for later analysis. The total meta-loss was averaged across tasks, and outer-loop optimization was performed by backpropagating the loss. Finally, the main model’s weights were updated using the adapted model, and the function re-turned the final loss along with stored query labels and predictions. This process helped the model learn a generalizable initialization that can quickly adapt to new tasks with minimal updates.
• Meta Validation: A meta-learning model was then evaluated by iterating over a set of tasks, each containing a support set for fine-tuning and a query set for evaluation. It first initialized the total loss and prepared storage for true labels and predictions. As it looped through tasks, it moved data to the appropriate device (CPU/GPU) and fi-ne-tuned a copy of the model on the support set using an inner-loop update function (inner update). After adaptation, it made predictions on the query set, applied a sig-moid activation function to convert logits into probabilities, and stored both true and predicted labels. Then, the function computed the binary cross-entropy loss for each task, accumulated the losses, and then averaged them over all tasks. Finally, it in-voked an early stopping mechanism, which can determine if the model should be saved or training should halt based on the meta-loss trend. Then the function re-turned the averaged loss, true query labels, predicted labels, and the early stopping decision, ensuring model evaluation in a meta-learning framework.

3.4.2. Meta Testing on Dataset

A new dataset, which the model had not seen during the meta-training phase, was prepared. This dataset contained new tasks from two different scenarios (restaurant and sidewalk) that the model needed to adapt to. Then, like the meta-training phase, tasks were sampled from the novel dataset. Each task was defined with a small number of training examples (support set) and testing examples (query set).

For each sampled task, the model adapted its parameters using the support set. This was usually done using the same adaptation procedure (inner loop) as in the meta-training phase. The key difference here was that the initial model parameters or meta-parameters, obtained from meta-training, were used as the starting point. After adaptation, the model was evaluated on the query set of the same task. The performance was recorded to assess how well the model had adapted to the new task.

3.5. Model Agnostic Meta-Learning

In the context of video anomaly detection, each task T_i corresponds to a different set of video frames, representing a unique scenario or environment. The collection of all such tasks is denoted by T = {T₁, T₂,…,T_n}. For each task T_i, there exists a dedicated dataset D_i = {(x_j,y_j)}mi for j = 1, where x_j is the input and y_j is the corresponding level. The model used for anomaly detection is parameterized by θ, which represents the set of all learnable parameters. During meta-training, the goal is to find an optimal initialization of θ such that, after a few gradient updates using data from any task T_i, the model quickly adapts and performs well.

3.5.1. Meta-Training Phase

The meta-training phase aims to find an initial set of model parameters, denoted as θ, that can rapidly adapt to new tasks. Within this phase, the inner loop involves task-specific adaptation: for each task T_i, the model performs a few steps of gradient descent to adapt the initial parameters θ, resulting in task-specific parameters θ′_i

θ′_i = θ − αΔ_θL_Ti(θ)

(9)

where α is the learning rate and L_Ti(θ) is the loss function for task T_i.

Following this, the outer loop performs a meta-update by adjusting the initial parameters θ based on the performance of the adapted parameters θ′_i across all tasks, enabling better generalization to unseen tasks.

θ<   −θ − βΔ_θ∑_iL_Ti(θ′_i)

(10)

where β is the meta-learning rate.

3.5.2. Meta-Testing Phase

During the meta-testing phase, the learned initial parameters θ are fine-tuned on a new dataset corresponding to a previously unseen task using a few gradient descent steps. Given a new task T_new with its dataset D_new, the model adapts its parameters according to the following update equation:

θ′_new = θ – αΔ_θL_Tnew(θ)

(11)

where α is the learning rate and L_Tnew(θ) represents the loss on the new task. This adaptation allows the model to quickly generalize new tasks using minimal training data.

3.5.3. Loss Function for Anomaly Detection

For anomaly detection, the loss function Lₜᵢ(θ) can be designed based on either the reconstruction error or the output of a discriminator. In a reconstruction-based approach, the loss function measures how well the model can reconstruct the input data, and it can be defined as:

L_Ti(θ) = Σ_(x,y)∈_Di‖x − f_θ(x)‖^²

(12)

where f_θ(x) is the reconstructed output of the model. This formulation penalizes large deviations between the input x and its reconstruction, thereby helping the model learn normal patterns and identify anomalies as instances with high reconstruction errors.

3.5.4. MAML Architecture

Figure 5 illustrates the core concept of MAML, where the objective is to find an optimal set of initial parameters θ that can quickly adapt to a variety of tasks with minimal training. The solid curve represents the trajectory of meta-learning, which optimizes the initial parameters across multiple tasks. Each red arrow (∇L₁, ∇L₂, ∇L₃) indicates the task-specific gradient computed during the inner loop of training for different tasks. These gradients guide the adaptation of the initial parameters θ toward task-specific optima (θ₁*, θ₂*, θ₃*), shown as blue dashed arrows labeled as “Learning/Adaptation”. The goal of MAML is to optimize θ such that a small number of gradient steps can lead to effective task-specific models, enabling rapid learning on new tasks. The figure highlights how meta-learning integrates feedback from multiple tasks to update the shared initialization in a way that supports efficient fine-tuning.

To implement the meta-learning approach in a task-agnostic manner, the following algorithm outlines the generic MAML procedure. This algorithm is designed to optimize an initial model that can quickly adapt to new video frame-based anomaly detection tasks with only a few gradient steps. The steps below detail the inner and outer learning loops required to achieve this fast adaptation (see Algorithm 1):

Algorithm 1: Generic MAML algorithm.

Require: D(T): Distribution over tasks (video frames) Require: α and β: step size hyper parameters Step 1. Randomly initialize parameter θ Step 2. While not done do Step 3. Sample batch of task Ti ∼ D(T) Step 4. For all Ti do Step 5. Evaluate ∇θLTi(fθ) with respect to k samples Step 6. Calculate adapted parameters using gradient descent: θ′i = θ-α∇θLTi(fθ) Step 7. End for Step 8. Update θ←θ − β∇θ∑Ti∼D(T)LTi(f′θ) Step 9. End while

4. Result and Analysis

In this study, MSAD videos were extracted into frames. Then, these frames were preprocessed by resizing to 224 × 224. Normalization using means and standard deviations was used.

After normalization, input images were given to swin transformer, which gave 768 feature vectors. These vectors were then given to detector models for classification into anomalous or normal. The output of the study was to classify input video frames into anomalous and normal types illustrated in Figure 6.

The experimental setup contained the given MAML parameters as illustrated in

Table 1. The MAML parameters and hyper parameters used in the implementation.

Metrics	Values
Epoch	100
K-Shot	10
K-Query	1
Optimizer	Adam
N-Way	5
Inner Learning Rate	0.1
Outer Learning Rate	0.01
Batchsz	2000
Input Image Size	224 × 224
No. of Layers	3
Feature Embedding Size	768
Device	Cuda

. Using these parameters, the model was trained, validated, and tested. A ten-shot one-query approach was implemented.

4.1. Confusion Matrix

The confusion matrix for model evaluation is given in Figure 7, where the matrix consists of four key values: true positives (1668), where the model correctly identified class “1”; true negatives (1671), where it correctly identified class “0”; false positives (329), where it mistakenly classified a “0” as “1; and false negatives (332), where it incorrectly classified a “1” as “0”. From these values, key performance metrics were calculated.

The performance matrix can be expressed in terms of accuracy, precision, recall, and F1-score, which are calculated as follows:

• Accuracy = (TP + TN)/(TP + TN + FP + FN) = 3339/4000 = 0.834
• Precision = TP/(TP + FP) = 1668/(1668 + 329) = 1668/1997 = 0.835
• Recall = TP/(TP + FN) = 1668/(1668 + 322) = 1668/2000 = 0.834
• F1-Score = 2 × (Precision × Recall)/(Precision + Recall) = 0.834

For video anomaly detection, this confusion matrix indicates that the model performs reasonably well, correctly classifying 83.4% of cases. The F1-score, a balance between precision and recall, is 83.4%, showing consistent performance. However, the presence of 332 false negatives suggests that some anomalies are being missed, which could be critical in applications where detecting anomalies is the priority. The false positives (329) indicate some normal events are being flagged as anomalies, which may lead to unnecessary alerts. Given the nature of anomaly detection, improving recall (currently 83.4%) is crucial to minimize missed anomalies. Fine-tuning the model, adjusting thresholds, or incorporating additional features (temporal) using BiLSTM or 3D swin model as a feature extractor could help reduce false negatives and improve overall detection performance.

4.2. Learning Curve

Figure 8 shows the two kinds of model evaluation, the Receiver Operating Characteristic (ROC) and the Precision-Recall (PR) curve. The ROC curve shown in Figure 8a, evaluates the performance of the video anomaly detection model by illustrating the trade-off between the true positive rate (TPR) and the false positive rate (FPR) at different classification thresholds. The red curve represents the actual model’s performance, while the blue dashed line serves as a baseline, indicating random guessing. A well-performing model should have its ROC curve significantly above this baseline, which is the case here. The steep initial rise of the curve suggests that the model can detect a substantial portion of anomalies with a low false positive rate, which is crucial for anomaly detection. However, as the curve flattens, increasing sensitivity results in a higher number of false positives, potentially leading to unnecessary alerts. The Precision-Recall (PR) curve shown in Figure 8b for video anomaly detection provides insight into the model’s ability to distinguish anomalies from normal events. Initially, at very low recall values, the model achieves extremely high precision, indicating that the first few detected anomalies are highly accurate with minimal false positives. As recall increases, meaning the model detects more anomalies, precision remains relatively stable before gradually declining. This suggests that while the model can correctly identify many anomalies with high precision, its performance degrades as it attempts to detect a larger number of anomalies. The sharp decline in precision at higher recall values indicates that an increasing number of normal events are being misclassified as anomalies, leading to a rise in false positives. This trade-off is common in anomaly detection systems, where optimizing high recall often results in reduced precision.

To improve the model’s effectiveness, techniques such as better feature extraction, threshold tuning, or post-processing methods like spatial smoothing can help reduce false positives while maintaining a balance between precision and recall. The optimal operating point for the model depends on the application—whether prioritizing precision to minimize false alarms or maximizing recall to ensure all potential anomalies are detected.

4.3. Model Comparison

Table 2. Model comparison.

Model	AUC
SAAD	0.69
RTFM	0.867
MGFN	0.85
UR-DMU	0.72
TEVAD	0.83
MLP + MAML + SWIN (our model)	0.91

represents the comparative analysis of six different models, including Scenario Adaptive Anomaly Detection (SAAD) [1,27], Robust Temporal Feature Magnitude (RTFM), Magnitude-Contrastive Glance and Focus Network (MGFN), Uncertainty Regulated Dual Memory Unit (UR-DMU), Text Empowered Video Anomaly Detection (TEVAD) [27], and the anomaly detector model (MLP + MAML + SWIN). The MLP + MAML + SWIN model outperformed all other models in comparison due to its unique combination of meta-learning and advanced feature extraction. Here, the SAAD model adopted a reconstruction-based few-shot learning strategy using U-Net and ConvLSTM to predict future frames and score anomalies via PSNR. RTFM and MGFN are representative weakly supervised methods that rely on pre-extracted features using I3D and swin and multi-instance learning-based objectives, focusing on identifying abnormal segments from video-level labels. Similarly, UR-DMU was designed to model normal behavior representations effectively while simultaneously extracting distinctive features from abnormal data, enabling a more accurate differentiation between normal and anomalous states. By integrating both visual and textual information, TEVAD enhances spatio-temporal features with semantic insights, allowing for more meaningful interpretation of abnormal events. All RTFM, MGFN, UR-DMU, and TEVAD models used swinT and I3D as their backbone architecture for anomaly detection.

Unlike traditional models like SAAD, which achieved a lower AUC of 0.69, or RFTM, MGFN, UR-DMU, and TEVAD, which reached AUCs of 0.867, 0.85, 0.72, and 0.83, respectively, the implemented model achieved a remarkable AUC of 0.91, where the anomaly detector model distinguished itself by combining swin transformer for robust spatial feature extraction with a meta-learned MLP classifier trained using MAML. In this model, the five-shot one-query approach was adopted for episodic training setup, and a ten-shot one-query approach was adopted for testing setup, enabling strong adaptation across diverse scenarios and camera views. This few-shot classification-based formulation allowed for efficient learning with limited examples and achieved superior AUC scores compared to all other evaluated methods, highlighting the strength of combining a powerful visual backbones with meta-learning in a classification-driven anomaly detection pipeline. This significant improvement is largely attributed to the integration of MAML, which allows the model to adapt quickly to new scenarios with minimal data, a crucial feature in video anomaly detection, where anomalies vary across different contexts. The swin transformer backbone enhanced feature extraction by effectively capturing spatial dependencies in the video frames through its attention mechanism. This enabled the model to accurately identify even subtle anomalies in complex video data. Furthermore, the combination of MLP with MAML and swin transformer leveraged the strengths of both powerful feature extraction and robust decision making, ensuring optimal performance.

5. Conclusions

In this study, a comprehensive approach to video anomaly detection was presented by leveraging a multi-scenario anomaly detection dataset containing diverse normal and anomalous video sequences. Frames were extracted from these videos, resized to 224 × 224 pixels with three color channels, and normalized to ensure consistency and optimal input formatting for the model. The proposed method demonstrated the effectiveness of the swin transformer in capturing detailed spatial dependencies within video frames, while a detector model utilizing a 10-shot one-query MAML approach was employed to classify frames as normal or anomalous. This combination of swin transformer for feature extraction and MAML for meta-learning resulted in significant improvements in anomaly detection accuracy and demonstrated superior generalization to novel and complex datasets.

However, several limitations remain: the current approach does not fully exploit temporal dynamics over long video sequences, which could hinder detection of anomalies evolving over time; the computational complexity of the swin transformer may limit real-time deployment, especially on resource-constrained devices; and despite the benefits of meta-learning, the model’s performance may degrade on entirely unseen anomaly types or low-quality data.

Future work could address these issues by integrating spatiotemporal models such as 3D swin transformers or temporal networks like BiLSTM to better capture sequential information, applying model compression techniques for efficiency, exploring unsupervised or self-supervised meta-learning paradigms to reduce reliance on labeled data, and validating the approach on additional diverse datasets to improve cross-domain robustness.

The findings of this research have important implications for both academic and practical fields, offering advancements for automated surveillance and public safety systems, smart city infrastructure monitoring, and anomaly detection in retail and industrial environments. Moreover, this work contributes a strong baseline for future studies investigating meta-learning and transformer-based architectures in video anomaly detection.