Approved by AICTE , Accredited by NBA , New Delhi & Affiliated to JNTU, Hyderabad
Bheemaram,Hasanparthy,Warangal,Telangana-506015 ,2024-2025
DEPARTMENT OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING
MINI PROJECT PRESENTATION
ON
SUSPICIOUS ACTIVITY DETECTION USING YOLOv11 and XGBOOST
Guidance by: Presented by:
N. Sandhyarani 22TK1A6644 Suddapally Stephenson
22TK1A6621 Neerella Harshavardhini
22TK1A6607 Bommineni keerthana
23TK5A6601 Yarva Akhil
CONTENTS
1.Abstract
2.Introduction
3.Architecture
4.Objectives
5.Background work
6.Methodology
7.Algorithm
8.Merits and Demerits
9.Result
10.Conclusion
11.References
ABSTRACT
The increasing need for intelligent surveillance has driven advancements in real-time
suspicious activity detection. This study integrates YOLO11 Pose Estimation with XGBoost to
develop an efficient and scalable security monitoring system. YOLO11 extracts skeletal
keypoints from individuals in video feeds, enabling precise movement analysis, while XGBoost
classifies behaviors such as loitering, sudden movements, and physical altercations as normal
or suspicious. A dataset of human activities will be used to train and evaluate the model,
incorporating feature selection, optimization techniques, and real-time factors like occlusions,
lighting variations, and crowd density. The study aims to enhance automated surveillance,
minimize false alarms, and reduce dependency on manual monitoring. The proposed system
will improve security response efficiency, optimize resource allocation, and contribute to safer
urban environments by enabling proactive threat detection.
Project Domain: Machine Learning, Deep Learning
Expected Outcome from the project work: The project aims to develop an AI-
driven model for real time activity classification, improving threat detection
accuracy, minimizing false positives, and supporting intelligent security solutions.
INTRODUCTION
The rise in security concerns has led to the need for intelligent surveillance systems.
Traditional monitoring relies on manual observation, which is time-consuming and
error-prone.
This project integrates YOLO11 Pose Estimation and XGBoost for real-time
suspicious activity detection.
The system enhances surveillance by automating threat detection and minimizing
false alarms.
ARCHITECTURE
CCTV / Video Feed |
|
v
Frame Extraction ← (e.g., 5 FPS)
|
V
YOLOv11 Detector ← Detects persons, bags, weapons, etc.
|
+---------V--+-------------+
| |
v v
-> Detected Objects -> Spatial-Temporal Features
-> (bounding boxes, Speed, direction,
classes, scores) duration, body posture
-> Person-object distance
+---------------------------+
|
V
Feature Aggregator
(Combines detection + time
|
V
XGBoost Classifier
(Suspicious vs Normal)
|
v
Suspicious Activity Alerts
(Timestamps, bounding boxes)
Objectives
Develop an AI-based surveillance system using YOLO11 and XGBoost.
Enhance real-time threat detection accuracy while reducing false positives.
Analyze human skeletal movements for suspicious behaviors like loitering,
sudden movements, and physical altercations.
Optimize system performance by addressing occlusions, lighting variations, and
crowd density challenges.
Improve response efficiency and resource allocation in security monitoring.
LITERACTURE SURVEY/BACKGROUND WORK
Traditional Surveillance Systems: Dependent on human operators, prone to
fatigue and missed detections.
Computer Vision in Security: Use of CNNs, RNNs, and deep learning for object
and activity recognition.
YOLO for Object Detection: High-speed real-time object detection widely used
in surveillance.
XGBoost in Classification: Efficient gradient-boosting algorithm for behavioral
analysis.
Existing Research: Studies on human activity recognition using pose estimation
and machine learning.
METHODOLOGY
Data Collection: Acquire datasets of human activities (normal and suspicious).
Preprocessing: Apply data augmentation, normalization, and noise reduction
techniques.
Feature Extraction: Use YOLO11 to extract skeletal keypoints and movement
patterns.
Behavior Classification: Train XGBoost to classify movements as normal or
suspicious.
Model Optimization: Fine-tune hyperparameters to improve accuracy and reduce
latency.
Testing & Evaluation: Validate the model against real-world scenarios.
Deployment: Integrate the model into a real-time security surveillance framework.
ALGORITHM
•Video Input: Surveillance video is fed into the system.
•Object Detection (YOLOv11): YOLOv11 processes each video frame, detecting
and locating objects (e.g., people, vehicles) and classifying them.
•Feature Extraction: The detected objects and their features (e.g., bounding box
dimensions, location, relationships with other objects) are extracted.
•XGBoost Classification: These features are used as input to the XGBoost
classifier, which predicts whether the detected activity is "normal" or
"suspicious".
•Alert Generation: Based on the XGBoost classification, the system can
generate alerts, such as sending notifications or triggering alarms, for
suspicious activities.
MERITS AND DEMERITS
Merits:
✔ Real-time Processing: Ensures instant detection and response.
✔ High Accuracy: YOLO11 and XGBoost together improve classification precision.
✔ Automation: Reduces the need for manual monitoring.
✔ Scalability: Cloud-based deployment supports large-scale applications.
Demerits:
⚠ Computational Cost: Real-time processing requires high-end hardware.
⚠ Data Dependence: Model performance relies on diverse and high-quality training datasets.
⚠ Privacy Concerns: AI-powered surveillance raises ethical and legal concerns.
RESULT
CONCLUSION
The proposed system enhances real-time security monitoring using YOLO11 and
XGBoost.
Improves threat detection accuracy, minimizes false positives, and optimizes
security resource allocation.
Addresses real-world challenges such as lighting variations, occlusions, and crowd
density.
The system contributes to proactive threat detection, making public spaces safer
and security monitoring more efficient.
REFERENCES
Redmon, J., & Farhadi, A. (2018). YOLO: Real-Time Object Detection. arXiv
preprint.
Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. ACM
Proceedings.
Zhang, J., & Tao, D. (2021). Human Activity Recognition in Video Surveillance.
IEEE Transactions.
Official YOLO Documentation: https://pjreddie.com/darknet/yolo/
Official XGBoost Documentation: https://xgboost.readthedocs.io/
