Table 1 Empirical review of existing methods.
References | Method used | Findings | Results | Limitations |
|---|---|---|---|---|
Accident detection using CNN | Accurate classification of traffic accidents | Achieved 89.5% accuracy in accident detection | Limited scalability to different accident types | |
Attention-based CNN-LSTM | Effective in violence detection with an attention mechanism | 85.3% accuracy on the UCF-Crime dataset | Limited to video data, no multimodal integration | |
Fuzzy Cognitive Deep Learning | Captures crowd emotions using cognitive models | 88% accuracy in predicting crowd behavior | Difficult to generalize to non-crowd scenarios | |
Multiple-scale motion consistency | Detects crowd-level abnormal behaviors | 92% AUC for crowd risk estimation | High computational complexity for large crowds | |
Deep Life Modeling for Crowd Monitoring | Accurate dynamic crowd modeling on edge devices | 90% accuracy for crowd monitoring | Sensitive to network latency in edge environments | |
Deep Graph Convolutional Neural Networks | Effective crowd characterization in structured and unstructured crowds | 87% F1-score in crowd classification | Ineffective in very sparse or low-density crowds | |
FSC-Set CNN for crowd counting | Accurate counting and localization of football crowds | 94% accuracy in crowd estimation | Focused only on sports environments, limiting generalization | |
GAN-based Crowd Management for Umrah | Real-time alert generation for crowd incidents | 91.7% accuracy in crowd incident prediction | Lack of generalizability to other religious events | |
VR-based crowd motion generation | Enables single-user crowd simulation in VR | 84.6% accuracy in generating realistic crowd motions | Limited scalability in generating large crowds | |
Ant Colony Optimization for fence layout | Optimized crowd management through fence layout | Reduced crowding by 23% in simulated environments | Only tested in simulation, not real-world data | |
Crowd descriptors for gathering understanding | Provides interpretable crowd gathering analysis | 88% accuracy in crowd density estimation | Lacks real-time applicability for large datasets | |
PublicVision Smart Surveillance System | Secure crowd behavior recognition using transformers | 90% accuracy in recognizing abnormal crowd behaviors | Limited scalability in larger urban areas | |
Attention-guided crowd counting | Improved crowd counting accuracy using segmentation-guided networks | 92.1% accuracy on large-scale datasets | Requires high computational resources | |
Variational Autoencoder with Motion Consistency | Detects abnormal crowd behaviors with variational models | Achieved 86% accuracy in detecting motion anomalies | Inefficient in low-resolution video scenarios | |
Transfer Learning for Crowd Emotion Prediction | Predicts human-vehicle interaction using crowd emotions | 89% accuracy in emotion-based anomaly detection | Limited emotion categories used in training | |
Fuzzy Decision Rules for Crowd Evacuation | Extracts decision rules for crowd evacuation strategies | 87% accuracy in predicting evacuation paths | Ineffective in non-crowded scenarios | |
Social Force Model for Crowd Evacuation | Models behavioral heterogeneity in crowd evacuations | 85.5% accuracy in evacuation simulation | Limited adaptability to different cultural behaviors | |
Emotional Contagion-Aware Reinforcement Learning | Simulates antagonistic crowd behavior with emotion contagion | 91% accuracy in modeling crowd emotions | High computational cost for large-scale simulations | |
CrowdGAN for video generation | Generates identity-free crowd videos using GANs | 84% realism score in video generation | Struggles with complex, high-density crowd environments | |
Anticipation modeling for crowd interaction | Models mutual anticipation in crowd behavior | 89% accuracy in predicting crowd interaction | Computationally expensive for real-time applications | |
Convolutional Recurrent Neural Networks | Forecasts citywide crowd transitions | 88.5% accuracy in predicting crowd flow | Limited to urban environments, not applicable in rural areas | |
IoT-based crowd flow prediction | Real-time urban crowd flow prediction | Achieved 92% accuracy in real-time crowd flow prediction | Limited IoT infrastructure scalability | |
Radar-based gait recognition | Recognizes crowd behavior using radar micro-Doppler signatures | 85% accuracy in open-set gait recognition | High sensor cost for large-scale deployments | |
Hidden Markov Model for gait detection | Detects abnormal gait patterns using vibration signals | 82% accuracy in detecting gait anomalies | Sensitive to sensor noise and interference | |
Deep One-Class Classifier for Parkinson’s patients | Predicts freezing of gait in Parkinson’s patients | 86% accuracy in predicting gait freezing | Limited to specific patient groups | |
Multimodal Emotion Recognition | Accurately recognizes emotions using situational knowledge | 89.2% accuracy in multimodal emotion detection | Limited by lack of contextual data for training | |
C3D for Crowd Behavior Detection | Detects crowd anomalies in large events like Hajj | 90% accuracy in action recognition | Struggles with extreme crowd densities | |
Meta-Heuristic Algorithm for Anomaly Detection | Detects anomalies in crowded environments | 87.5% accuracy in identifying public safety risks | Lacks real-time capability for large-scale crowds | |
Graph Convolutional Neural Networks | Detects abnormal crowd behavior using GCN | 85% accuracy in detecting graph-based anomalies | Sensitive to incomplete graph data | |
Transfer Learning for Suspicious Crowd Behavior | Detects suspicious human crowd behavior | Achieved 89% accuracy in anomaly detection | Requires large pre-trained models for deployment | |
Pre-Trained CNN for Crowd Anomaly Detection | Efficient anomaly detection in crowd videos | 87% accuracy in detecting violent actions | Limited generalization to non-violent behaviors | |
Statistical Physics for Behavior Detection | Models crowd behavior using entropy-based models | 85.6% accuracy in detecting abnormal crowd dynamics | Computationally expensive for real-time processing | |
Temporal Association Rules for Crowd Modeling | Models crowd behavior using temporal rules | 83% accuracy in predicting crowd transitions | Limited to well-structured crowds | |
Hybrid Neural Networks for Behavior Detection | Detects abnormal human behavior in crowded scenes | 88% accuracy in identifying suspicious behaviors | Limited adaptability to varying crowd sizes | |
Zero-Shot Classifier for Anomaly Detection | Detects anomalies using spatio-temporal descriptors | 85% accuracy in detecting novel anomalies | Sensitive to inaccurate descriptors | |
GeoVideo for Regional Crowd Analysis | Analyzes regional crowd status using multimedia data | 86% accuracy in crowd quantity estimation | Requires high-quality social media data integration | |
Deep Learning for CCTV Surveillance | Generates real-time alerts in CCTV surveillance | 90.3% accuracy in real-time anomaly detection | Requires large computing infrastructure for real-time use | |
YOLO + Conv2D Net for Abnormality Detection | Detects abnormal human behavior in real-time | 88.5% accuracy in human behavior recognition | Inefficient in high-latency environments | |
Congestion-Aware Path Planning | Plans paths considering spatial-temporal crowd anomalies | 85% accuracy in congestion-aware path planning | Sensitive to sudden crowd dynamics changes | |
GANs for Dynamic Image Representation | Detects crowd anomalies using dynamic image representations | 87% accuracy in detecting optical flow anomalies | Struggles with high-resolution image data |
