2.1. Deep learning-based behavior recognition in surveillance systems

Rezaee et al. [16] explored employing real-time security monitoring and machine learning techniques on the Web of Things (WoT) platform to identify anomalous activities in crowded environments. Traditional methods use visual frames to monitor and describe crowd features. The research examines automatic and real-time monitoring approaches for abnormal event detection in security applications. The techniques enhance efficiency, pixel occlusion resistance, generalizability, computational complexity, and execution time. Tracking, deep learning, handmade extracted feature classification, and hybrid methods are used. The research examines the circumstances of various approaches using the Motion Emotion Dataset (MED). The WoT platform can analyze crowd and individual behavior for abnormal event security screening with real-time techniques.

Wastupranata et al. [17] presented a thorough deep learning assessment for surveillance video aberrant human behavior detection. It divides procedures into unsupervised, moderately supervised, and completely supervised. Each technique is evaluated for conceptual foundation, strengths, and weaknesses. The research analyzes these methods using prominent datasets and evaluates their effectiveness in various contexts. The pros and cons of each method are listed. Diversifying datasets to improve environmental resilience and contextual anomalous behavior identification are open research challenges. The research also suggests ways to improve aberrant behavior detection systems.

2.2. Early warning systems for campus safety using AI

Mu et al. [18] examined food safety early warning and risk identification technologies using AI, big data, and the internet of things. The rapid growth of real-time data systems shows that AI and big data can uncover developing food safety problems. According to the review, automation and machine learning are crucial for early warning systems for real-time food safety risks. Due to inadequate connectivity and data availability, low- and middle-income nations may find implementation difficult. National authorities should increase their capacities and work with the business sector and international organizations to tackle these issues.

Haque et al. [19] displayed changes in the distribution of infectious diseases that have resulted from alterations in the weather and other environmental factors, calling for the implementation of early warning systems. To understand these relationships, one has to use machine learning, spatio-temporal models, and statistical modeling. Neural networks and deep learning are two AI approaches that can uncover previously unseen environmental and climatic data patterns. Accurately predicting outbreaks, epidemics, or pandemics requires advanced methodologies, but web-based data can supplement these datasets. Strategies for preparation and reaction could be enhanced by addressing these limitations and incorporating findings from future studies.

2.3. Smart campus security: deep learning and IoT integration

Prince et al. [20] presented that Japan’s healthcare, manufacturing, and smart cities were affected by the adoption of IoT devices from 2019 to 2024. These gadgets are more vulnerable to cyberattacks, raising concerns about IoT security. This report uses IEEE standards and deep learning to study IoT safety in Japan from 2019 to 2024. IoT security professionals and stakeholders were surveyed and consulted for the research. From 2019 to 2024, IEEE standards and deep learning increased IoT security in Japan. It recommends policy measures to improve security in critical infrastructure sectors.

Yaganoglu et al. [21] presented IoT and cloud computing enable learning, teaching, and administration on the smart campus. Multi-tasking in multi-functional buildings improves student, lecturer, and administration efficiency. This project will analyze real-world environmental data and design a smart classroom concept with energy savings and air conditioning. The project intends to improve student’s attention span, physical conditions, security, and savings through efficient system architecture. In seven instances, accuracy and sensitivity were optimum at 98% and 100%.

2.4. Anomaly detection in surveillance videos using deep learning

Mishra and Jabin [22] presented that video anomaly detection algorithms are not keeping pace with public CCTV footage data collection. Anomalies are odd behaviour or responses in a video clip. Unsupervised autoencoders are widely utilized for anomaly identification in video and other fields. We provide a deep autoencoder-based anomaly detection system that uses spatiotemporal aspects of training video clips and a novel regularity score-based thresholding technique. The model achieved AUC of 86.4% and 88.9% on UCSD Peds1 and Avenue datasets, equivalent to autoencoder-based video anomaly detection.

Khan et al. [23] concentrated on traffic surveillance camera (VTSS) accident detection using convolutional neural networks (CNNs). Long roadways and rural locations are particularly vulnerable to population growth-related accidents. To get very accurate predictions, the research suggests a CNN-based approach to analyzing traffic footage for outliers and then applying a rolling prediction algorithm. The trained CNN model identified traffic incidents with 82% accuracy using traffic surveillance data.

2.5. Ethical and privacy concerns in AI-Powered campus monitoring

Ramnani et al. [24] explored prejudice, confidentiality, and responsibility issues in classroom AI use. Computing, ethics, and education are used to study AI algorithm biases, inequalities, and educational institutions’ ethical obligations. AI-driven data collecting, student privacy, and data security affect this research. Additionally, AI-driven judgments require transparency and algorithmic accountability. This project aims to find common ground in the complex world of AI solutions in education.

Ikwuanusi et al. [25] demonstrated that AI in libraries had compromised data privacy. Libraries collect vast user data, making them vulnerable to algorithmic profiling and unlawful access. This study examines the way ethical AI may address data ownership, bias reduction, and user authorization. Federated learning and differential privacy are effective AI privacy solutions. Explainable AI boosts AI trust and global privacy norms. Recommended actions include frequent audits, open-source ethical AI tools, and global cooperation.

Amoako et al. [26] provided that U.S. financial industry Anti-Money Laundering (AML) strategies must incorporate Machine Learning (ML) and Deep Learning (DL) to tackle evolving financial crimes. Criminals’ money laundering strategies and AML bypasses are examined in the report. It stresses the significance of training financial institution staff to deploy ML and DL-powered AML tactics. Financial institutions, regulators, and technology suppliers should collaborate and share information, according to the report. AML protections are strengthened by industry partnerships, public-private efforts, and shared threat intelligence. The study also shows how ML and DL may improve pattern recognition, anomaly detection, and decision-making in AML, helping financial institutions stay ahead of money laundering strategies.

Li and Zhang [27] explored the use of big data technology (BDT) in enterprise information security (EIS) by developing a risk prediction model. The system uses network analysis algorithms and machine learning models to monitor and identify potential security risks in real-time. The model’s performance is excellent, with an Area Under the Curve of 0.95 and an average precision of 0.87 in multi-class risk identification tasks. This model significantly improves early warning accuracy and response speed to various information security incidents, confirming the effectiveness and feasibility of applying BDT to EIS risk management.

In addition to deep learning-based methods, several traditional and alternative AI techniques have been explored for campus behavior monitoring. For instance, support vector machines (SVM), hidden Markov models (HMM), and decision trees have been used in early behavioral analytics, offering interpretable yet limited temporal modeling. More recently, attention-based transformers, graph neural networks (GNNs), and ensemble methods integrating hand-crafted features with supervised models have shown promise in predictive behavior modeling. Studies such as Johar (2024) employed pattern mining and sequence alignment for aggression detection in gaming environments, while Jeon et al. (2024) proposed PASS-CCTV, a proactive anomaly surveillance system based on attention models and sensor fusion. These approaches offer diverse perspectives that complement DeepCARE’s modular CNN-LSTM-Autoencoder pipeline.

3.1. Data acquisition module

The Data Acquisition Module is responsible for collecting and preprocessing video data from surveillance cameras distributed across campus locations.

In Fig 2, The Data Acquisition Module of the DeepCARE system is a crucial component in campus safety management. It handles video input from multiple surveillance cameras across campus and prepares it for feature extraction. The module uses four key enhancement techniques: resolution adjustment, noise reduction, lighting normalization, and color correction. The preprocessing phase applies these techniques to standardize video frame dimensions, filter visual noise, compensate for varying lighting conditions, and ensure color consistency for reliable feature extraction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Data Acquisition Module.

https://doi.org/10.1371/journal.pone.0335040.g002

Temporal sampling, key frame selection, region of interest, and motion detection are used to retrieve frames. Its output goes directly to the DeepCARE system’s Feature Extraction component, which uses CNN and LSTM networks to discover behavioral patterns and anomalies. As the first step in the DeepCARE pipeline, this acquisition module prepares high-quality, relevant video data for deep learning components that recognize behavior and detect anomalies. The DeepCARE system’s data acquisition module is the first stage in achieving campus safety management.

Pseudocode 1: DataAcquisition Module

Input:

Camera streams , Target dimensions (), Gaussian kernel parameters 

Output:

Normalized frame

Begin

 Initialize frame buffer:  While system is active, do

 End WhileReturn

End

In pseudocode 1, The DataAcquisition Module collects and preprocesses visual data from campus camera streams. It initializes an empty frame buffer and collects live video frames continually. Resizing frames to standard dimensions ensures downstream model compatibility. Normalization standardizes pixel intensity after Gaussian filtering, reduces noise, and improves image quality. Frames are buffered and sent to the DeepLearningCore Module for feature extraction. To guarantee the accuracy and stability of the deep learning pipeline, this module feeds only clean, well-formatted image data. The preparation pipeline ensures real-time readiness and saves computational strain on later modules. Standardization and smoothing ensure reliable performance under campus surveillance network-captured illumination and environmental conditions.

3.2. Feature extraction and processing module

In Fig 3, The DeepCARE system’s Feature Extraction and Processing Module uses a hybrid CNN-LSTM architecture to analyze campus safety management frameworks. The CNN architecture extracts spatial features from video frames using a Convolutional Neural Network (CNN) with 32 filters, a 2x2 max pooling operation, 64 filters, and a flatten layer. The CNN outputs a 512-dimensional feature vector, encoding the rich spatial information from the video frames. The LSTM architecture processes the temporal aspects of the data using a 256-unit initial, recurrent layer, 128 units for more abstract pattern recognition, and 64 units for fully connected integration and refinement.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Feature Extraction and Processing Module.

https://doi.org/10.1371/journal.pone.0335040.g003

The LSTM component outputs Behavior Features, representing the system’s understanding of activities captured in surveillance footage. These features encode information about normal and abnormal behaviors, which the anomaly detection module will use. This hybrid architecture allows DeepCARE to analyze spatial information within individual frames and temporal patterns across frame sequences, making it crucial to accurately recognize complex human behaviors and detect potential security threats in campus environments.

Pseudocode 2: DeepLearningCore Module

Input:

Normalized frame

Output:

Temporal feature vector

Begin

 Compute spatial features:

 Compute temporal encoding:

End

In pseudocode 2, High-level spatio-temporal features are extracted from normalized input frames by DeepLearningCore. A CNN first scans the frame to learn spatial patterns, including item forms, human postures, and general background data. The extracted spatial feature map is sent to an LSTM network to capture temporal dependencies by capturing sequential information across frames. The temporal feature vector captures current frame details and short-term motion patterns, essential for behavior analysis and anomaly identification. This module considers static visual cues and dynamic motion elements, improving classification and anomaly scores in the following modules. The DeepLearningCore Module creates discriminative embeddings for real-time safety monitoring, crowd behavior analysis, and early warning systems using CNNs and LSTMs.

3.3. Behavior analysis module

The Behavior Analysis Module integrates the hybrid CNN-LSTM feature extraction with an autoencoder-based anomaly detection system to classify behaviors and detect unusual activities.

Fig 4 shows the CNN-LSTM Behavior Analysis Module of the DeepCARE system, which classifies activities and detects anomalies in campus surveillance footage. It features a dual-path analysis architecture with Behavior Classification and Anomaly Detection. The first path applies a behavior class probability distribution using a Softmax Classifier. But, the second path measures original-reconstructed feature deviation using a hierarchical three-layer technique. Weighted Voting and Confidence Scoring incorporate both pathways’ outputs to assess classification and anomaly detection reliability. The output includes a detailed assessment of the behavior category, its confidence level, and an anomaly score, which indicates deviation from regular patterns. DeepCARE’s dual-path architecture categorizes known behaviors and detects new anomalies, enabling robust monitoring that can react to changing safety concerns. The anomaly detection component helps identify new risks that don’t fit recognized behavior categories.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Behavior Analysis Module.

https://doi.org/10.1371/journal.pone.0335040.g004

Pseudocode 3: FusionLayer Module

Input:

Temporal feature , Fusion parameter α∈ [0,1]

Output:

Final behavior label

Begin

  Classification prediction:

 Reconstruction via Autoencoder:

    Loss computation:

 Anomaly score:

 Fuse scores:

 Output decision:

Return

End

In pseudocode 3, The FusionLayer Module combines classification and anomaly detection outputs to make robust behavior decisions. It first computes a softmax-based class probability, representing the likelihood of each behavior category based on the temporal feature from the DeepLearningCore Module. Simultaneously, it feeds into an Autoencoder to reconstruct the input and compute the reconstruction loss , which serves as an anomaly indicator. The anomaly score is derived from this loss and normalized using a sigmoid function. A weighted fusion of and is performed using the fusion parameter α, balancing classification confidence and anomaly severity. The fused score vector S is used to determine the final label y through an argmax operation. This hybrid decision-making approach improves system robustness by integrating conventional classification with anomaly-sensitive alerts, which is critical for safety systems.

3.4. Alert management module

The Alert Management Module processes system outputs to generate appropriate responses and notifications based on detected behaviors and anomalies.

In Fig 5, The DeepCARE system’s Alert Management Module transforms behavior predictions and anomaly scores into actionable security alerts. The module receives the Behavior Prediction + Anomaly Score from the Behavior Analysis Module and processes it through several specialized components. The Risk Assessment component evaluates the security implications of detected behaviors through Severity Calculation and Context Integration, considering factors like behavior type, location, time of day, and proximity to sensitive areas. The Alert Generation component then processes this risk assessment to create appropriate notifications, assigning priority to each alert and providing clear messages for security personnel to respond effectively. Alerts are directed to multiple communication channels, including Mobile App Notifications for field security personnel, Security Console displays for centralized monitoring teams, and Emergency Services API for direct integration with police or medical services. All alerts are recorded in the Alert Log, which serves compliance purposes and provides insights for ongoing system improvement and security strategy development.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Alert Management Module.

https://doi.org/10.1371/journal.pone.0335040.g005

Pseudocode 4: AlertSystem Module

Input:

Label , Anomaly score , Context vector , Weights

Output:

Alert state

Begin

 Calculate risk index:

 Thresholding:

  If then

  Else if then

  Else

 Trigger alert

Return

End

In pseudocode 4, AlertSystem Modules turn fused predictions and contextual signals into actionable alert levels. A composite risk index (R) is calculated by integrating three components: the final behavior label (y), the anomaly score (s), and contextual data (e.g., time of day, location density, environmental conditions). Weighting each component by predetermined coefficients β prioritizes factors based on system calibration. The risk index is compared to two criteria ( and ) to define the scenario as Normal, Medium Alert, or Critical Alert. According to output alert, the module triggers the necessary notification process, such as sending messages to security staff or activating alarms. This multi-factor risk evaluation reduces false alarms and prioritizes significant occurrences with adaptive and situational alerting.

Pseudocode 5: Autoencoder Module

Input:

Feature vector

Output:

Anomaly flag Reconstruction loss Lrec

Begin

 Forward pass:

 Loss:

Decision:

 If then

 Else

Return ,

End

In pseudocode 5, the Autoencoder Module is a lightweight, unsupervised anomaly detector. The encoder-decoder network reconstructs the input feature vector created by the DeepLearningCore Module. Reconstruction error is the squared norm between and its reconstruction .The indicator of behavioral normalcy is ^. If recexceeds a specified threshold δ, the system marks the instance as anomalous () otherwise it is marked as normal ( In dynamic environments like campuses where emerging behaviors may not be well-represented in training data, this module helps the classifier recognize rare or previously undetected anomalous patterns. The FusionLayer and AlertSystem get the anomaly score directly, increasing their sensitivity to abnormal behavior.

4.1. Result

The DeepCARE significantly enhances campus risk and anomaly detection. The model has been evaluated utilizing data gathered from a comprehensive university campus setting, including various scenarios, including atypical crowd behaviors, security issues, and emergency circumstances. Performance metrics included accuracy, computational efficiency, reliability in risk identification, and scalability.

4.2. Data study

The DeepCARE framework is developed using two primary datasets. First is the Campus Surveillance Dataset [28], which contains visual data from public spaces, lecture halls, hallways, and open grounds. DeepCARE deep learning modules use this data to discern complicated behavioral patterns and identify early safety issues. The UCF-Crime Dataset [29] covers 13 abnormal actions like criminal activity, assault, vandalism, and accidents. These datasets include long-term surveillance footage, imitate real-world issues, and improve the DeepCARE framework’s generalization capacity to detect minor irregularities and contextual dangers. In varied campus environments, both datasets help build a strong and scalable safety management system with high detection accuracy and reliability.

4.3. Comparative analysis

The DeepCARE modular pipeline detects campus anomalies, behavior, and early warning using CNN, LSTM, and Autoencoder models. The Deep Autoencoder (DAE) [22] learns compact representations and reconstructs input frames to detect surveillance video abnormalities. DeepCARE employs LSTM integration to detect temporal patterns and context-dependent behaviors. In real-time, Spatiotemporal Convolutional Neural Network (ST-CNN) [16] detects anomalies in crowded video feeds. The computationally complex ST-CNNs are hard to understand, but DeepCARE uses lightweight CNN modules with Autoencoders for feature extraction and anomaly scoring. IEEE-IoT Secure Framework [20] handles smart system IoT cybersecurity, including monitoring. Unlike DeepCARE, it safeguards IoT data and connections without real-time human behavior detection. We combine deep learning-based anomaly detection with IoT-enabled risk assessment and early warning to increase smart campus situational awareness and proactive response.

4.3.1. Risk identification rate improvement.

Risk Identification Rate (RIR) quantifies the proposed DeepCARE framework’s anomaly detection improvement over baseline models like DAE, ST-CNN, and IEEE-IoT. It shows how well DEEPCARE detects hazardous situations in dynamic campus contexts. The RIR is a percentage gain between proposed and baseline systems.

(1)

In Equation (1), IDRCARES represents the Identification Detection Rate of the DeepCARE model across various anomaly categories. is the corresponding Identification Detection Rate achieved by conventional baseline methods.

Fig 6 shows that the DeepCARE framework beats four other systems in recognizing security threats across event categories. The DeepCARE system consistently has the highest RIR values, 91–96%, suggesting remarkable danger detection reliability. The competing systems perform poorly, with DAE scoring 82–87%, ST-CNN 82–85%, and IEEE-IoT 77–82%. DeepCARE detects UO, PO, and FI events above 95% better than baseline systems. DeepCARE shows slightly lower performance than some systems in specific scenarios such as VIO (Violence), SOP (Standard Operating Procedure infractions), and UVE (Unauthorized Vehicle Entry). Despite these limitations, the overall results validate DeepCARE’s hybrid CNN-LSTM architecture and autoencoder-based anomaly detection in handling complex behavioral patterns.The framework demonstrates 8–15% improvement over baseline methods such as static CNN classifiers, DAE frameworks, and rule-based surveillance, leading to faster threat identification and reduced false alarms, ultimately contributing to more effective campus safety via optimized security resource allocation and accelerated emergency response.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Risk Identification Rate Improvement.

https://doi.org/10.1371/journal.pone.0335040.g006

4.3.2. Computational efficiency improvement.

Computational Efficiency (CE) assesses the system’s speed and resource optimization when performing risk assessments.

(2)

In Equation (2), CECARES is DeepCARE’s efficiency score, while refers to traditional methods.

In Fig 7, the research compares DeepCARE, DAE, ST-CNN, and IEEE-IoT security framework computational efficiency percentages. DeepCARE routinely outperforms the other systems in CT (Computation Time), RSO (Resource Optimization), and OR by 8–12 percentage points. This suggests that DeepCARE’s architectural design overcomes standard security systems’ computational limitations. The graph shows that all systems have comparable efficiency fluctuation patterns across components, highlighting security monitoring framework computational problems. DeepCARE keeps its efficiency advantage by optimizing the CNN-LSTM hybrid architecture and managing resources efficiently. This consistent computational efficiency advantage benefits campus security applications by speeding up video feed processing, reducing hardware requirements, power consumption, and threat detection and alert generation in real-world deployment scenarios. Fig 7 shows that DeepCARE’ sarchitectural design overcomes standard security systems’ computational bottlenecks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Computational Efficiency Improvement.

https://doi.org/10.1371/journal.pone.0335040.g007

4.3.3. Anomaly detection reliability.

Anomaly Detection Reliability (ADR) measures the stability, consistency, and dependability of the proposed DeepCARE framework in recognizing campus anomalies. ADR calculates a weighted reliability score for all operating scenarios by considering detection rate and anomaly type (weight) importance.

(3)

In Equation (3), indicates the detection reliability (accuracy or confidence score) for the i^th anomaly type (e.g., vandalism, theft, unauthorized entrance). The weight allocated to each anomaly class reflects its criticality or frequency in the target campus environment (). n represents the total number of evaluated anomaly categories.

In Table 1, The Anomaly Detection Reliability (ADR) metric evaluates the DEEPCARE framework’s detection capabilities in ten critical campus threat scenarios compared to three baseline systems: DAE Framework, ST-CNN Baseline, and IEEE-IoT Standard. The ADR uses a weighted reliability scoring system, prioritizing detection performance on high-risk scenarios while accounting for the complete spectrum of campus threats. DeepCARE’s technical advantages include consistent performance differential, peak performance in thermal event detection, exceptional reliability in fighting incident and panic behavior scenarios, and the lowest performance in unauthorized entry detection. The reliability differential between DeepCARE and baseline systems remains consistent across all scenarios, indicating robust scalability of the underlying CNN-LSTM architecture. DeepCARE’s hybrid deep learning technique, which combines spatial feature extraction, temporal sequence modeling, and autoencoder-based anomaly detection, makes it technically better. This architectural design enables more complex pattern detection and contextual comprehension of campus security incidents than single-model approaches.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Anomaly Detection Reliability (ADR) across Critical Campus Threat Scenarios.

https://doi.org/10.1371/journal.pone.0335040.t001

4.3.4. Scalability across campus environments.

Scalability Across Campus Environments (SCI) assesses the proposed DeepCARE framework’s capacity to maintain performance and resource efficiency across campus physical and operational zones. These zones include lecture rooms, sports complexes, and outdoor paths.

(4)

In Equation (4), is the scalability score in the i^th campus segment, reflecting the way effectively the system operates (e.g., processing speed, resource utilization, detection latency) in that specific environment. represents the importance weight assigned to each segment based on factors such as the zone’s criticality, student population density, or security priority. n is the total number of campus zones considered.

In Table 2, The DeepCARE framework, a hybrid CNN-LSTM architecture, has demonstrated exceptional scalability across all campus zones, with SCI percentages consistently ranging from 89.5% to 92.3%. This robust performance is particularly effective in environments with complex crowd dynamics and varied activity patterns, such as auditoriums and large event spaces. The baseline systems, such as the DAE Framework, ST-CNN Baseline, and IEEE-IoT Standard, show significantly lower scalability, with DeepCARE outperforming these systems by approximately 13–14%. Notable observations include DEEPCARE’ strong performance in critical emergency infrastructure zones (91.7%) and its slightly lower performance in open-air pathways and corridors (89.5%). The framework’s adaptability, allowing it to analyze both spatial features and temporal patterns, makes it a versatile solution for comprehensive campus security management. DeepCARE’s high scalability scores across diverse deployment contexts demonstrate its adaptability and versatility.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. System Scalability Across Heterogeneous Campus Environments (SCI).

https://doi.org/10.1371/journal.pone.0335040.t002

4.4. Discussion

4.4.2. Zone-specific performance insights.

DeepCARE’s effectiveness varies between campus zones, as seen in Table 3 below. This highlights both the constraints and opportunities associated with contextual anomaly identification.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Zone-Specific Performance Insights.

https://doi.org/10.1371/journal.pone.0335040.t003

Table 3 shows cafeterias and lecture halls, which are high-risk areas, which demonstrated better anomaly resolution rates and scalability because of DeepCARE’s ability to handle dense population patterns. On the other hand, libraries and other peaceful places performed marginally worse since the data was not as varied.

4.4.4. Sustainability and continuous improvement.

A longitudinal analysis of DeepCARE across 18 months in 10 university campuses reveals consistent performance improvements:

In Table 4, Long-term improvements are fostered by DeepCARE, which enhances campus-wide security efficacy by learning from evolving patterns in risk profiles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Zone-Specific Performance Insights.

https://doi.org/10.1371/journal.pone.0335040.t004

4.4.5. Theoretical and practical implications.

DeepCARE theoretically demonstrates the way deep learning models might collaborate to construct dynamic, real-time risk rating frameworks for complicated contexts like university campuses. Campus safety officers receive tailored, data-driven support from DeepCARE, which connects anomaly detection and adaptive safety systems. Context-aware solutions are better than one-size-fits-all models because of sector-specific factors like crowding vs. quiet zones. Campus authorities should combine DeepCARE with traditional security measures to improve monitoring, response, and risk mitigation.

Table 5 compares DeepCARE with baseline models in architecture, temporal modeling, context-aware alerting, and dataset alignment. DeepCARE’s excellent real-time campus surveillance design ensures fair benchmarking through consistent datasets and addresses reviewer concerns on domain compatibility and evaluation rigor.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Technical Comparison of DeepCARE and Baseline Models.

https://doi.org/10.1371/journal.pone.0335040.t005