REGIONAL INSTITUTE OF SCIENCE & TECHNOLOGY,

MEGHALAYA
Run by: ERD Foundation - Guwahati
Approved by AICTE, New Delhi & Affiliated to NEHU - Shillong

3rd International Conference

Automated Emergency Medical Vehicle (EMV) Dispatch System

Based on Severity Levels

Guided by
Mr. Karjan Basumatary
Registrar, RIST
Presented by

Bikram Tripathi 21RISTCSE004

Toni Deori 21RISTCSE035
Zahra Shaikh 21RISTCSE038

1
 CONTENT
1. INTRODUCTION…………………………………………………….3
2. MOTIVATION………………………………………………………...4-5
3. PROBLEM STATEMENT…………………………………………….6
4. RESEARCH GAP……………………………………………………..7
5. FEASIBILITY AND VIABILITY…………………………………….8-9
6. DESIGN……………………………………………………………….10-13
7. OBJECTIVES…………………………………………………………14
8. METHODOLOGY………...……..……………………………………15-28
9. TECHNICAL APPROACH………………...…………………………29-30
10. EARLY STAGES RESULT……..…………………………………….31-34
11. IMPACT AND BENEFITS………………...……………………..…..35-36
12. REFERENCES………………………………...………………………37-382
 INTRODUCTION
• Rapid medical intervention is crucial in reducing fatalities and severe injuries following
accidents. However, current EMV dispatch systems often face delays, resulting in worsened
outcomes for accident victims.
• A factor that contributes to the high number of deaths are delays in seeking medical treatment.
This refers to the time taken to send an injured victim to the emergency room being too long, due
to problems including delayed accident reporting, erroneously-reported accident locations, and
misjudgment of accident severity.
• This project aims to develop an automated system that uses CCTV footage to detect accidents in
real-time. The system will leverage deep learning to accurately identify incidents and assess their
severity based on visual data.
• Unlike traditional methods, this system prioritizes EMV dispatch based on the severity of the
accident. This ensures that serious cases requiring immediate attention are addressed swiftly,
improving both response times and resource management.
3
 MOTIVATION
• WHO reports over 1.3 million road deaths annually [1], highlighting the urgent need
for improved emergency response systems. This project directly addresses this
global issue by focusing on faster and more effective dispatch.
• Delayed emergency responses result in unnecessary loss of life. This system offers
peace of mind, ensuring that severe accidents receive immediate attention,
potentially saving countless lives.
• A 1-minute reduction in ambulance response time can boost survival rates by
24%[2]. This automated system uses CCTV to prioritize severe cases, improving
response efficiency and saving critical time.
• With growing urbanization and traffic, now is the perfect moment to leverage AI
and CCTV technology for real-time, severity-based dispatch, making emergency
responses more efficient and timely.
4
Year Total No. of Road Accidents Total No. of Fatalities Total No. of Persons Injured

2018 4,70,403 1,57,593 4,64,715

2019 4,56,959 1,58,984 4,49,360

2020 3,72,181 1,38,383 3,46,747

2021 4,12,432 1,53,972 3,84,448

2022 4,61,312 1,68,491 4,43,366

Fig. Total number of road accidents, fatalities and persons injured in the country from calendar year 2018 to
2022, data published by Ministry of Road Transport & Highways
 PROBLEM STATEMENT
The manual EMV dispatch system currently in use is plagued by several issues, including:
• Existing EMV dispatch systems often rely on manual reporting, which leads to slow
detection of accidents and delayed dispatch of emergency vehicles. This results in a
significant number of fatalities and worsened outcomes for victims due to late medical
intervention.
• Current systems dispatch EMVs without factoring in the severity of accidents, leading
to situations where ambulances are sent to minor incidents while severe cases face
longer wait times, increasing the risk of death or long-term injury.
• Manual reporting systems are inefficient in providing real-time, accurate assessments of
accident severity, resulting in suboptimal prioritization of EMV dispatch. A more
advanced system is needed to automate the detection and classification of accidents
based on their urgency.
6
 RESEARCH GAP
• YOLO Localization Issues: YOLO is fast but struggles with overlapping bounding
boxes, reducing detection accuracy in complex scenes like accidents.
• CNN's Speed-Accuracy Trade-off: CNNs are more accurate but too slow for real-time
use, making them impractical for emergency dispatch scenarios.
• Temporal Dynamics: Most models fail to capture the motion over a sequence of frames,
missing critical accident events.
• Occlusions & Complex Motions: Current methods misinterpret occlusions and complex
movements, leading to false positives or missed detections.
FEASIBILITY AND VIABILITY

8
 Challenges Feasible Solutions
Accurate Detection in Use YOLO’s object detection alongside Optical Flow to improve detection in
Complex Traffic crowded or visually complex environments.
Environments ghghghghghghghghhhhhhhhhhhhhh
Temporal Dependencies Implement LSTM networks to analyze sequences of events and classify
in Severity severity more accurately, based on prior data and real-time input.
Classification fhfhfhfhfhfhfhffhhffhhh
Real-Time Processing Utilize edge computing and scalable cloud infrastructure to manage and process
of Large Video Feeds large volumes of CCTV footage in real time.
Limited Data for Model Continuously update models using new accident footage and simulated data to
Training enhance accuracy over time.
Integration with Develop APIs and interfaces for smooth integration with current dispatch
Existing Dispatch centers and medical systems, minimizing operational disruptions.
Systems hhhhhhhhhhhh
Cost of Implementation Leverage existing public and private CCTV networks to reduce infrastructure
hhh costs while securing government and private partnerships for funding.
DESIGN
BLOCK DIAGRAM

10
Figure 1: Block Diagram

11
 DESIGN
USE CASE DIAGRAM

12
Figure 2: Use Case Diagram

13
 OBJECTIVES
1. The primary objective is to collect and curate real-world datasets containing images of
traffic scenarios and accidents from publicly available sources, such as YouTube, COCO,
and Roboflow. This dataset will form the foundation for training and evaluating the
performance of various deep learning models.
2. The secondary objective is to conduct a comprehensive comparative analysis of different
deep learning algorithms for image detection, including YOLO (versions 5 and 8). This
analysis will evaluate the accuracy, speed, and performance of each algorithm in detecting
accidents and other relevant objects in traffic scenes.
3. The third objective is to develop and implement a hybrid model combining YOLO,
Optical Flow and LSTM.

14
METHODOLOGY

15
1. Dataset Collection and Preparation
STEP 1. Data Sourcing- Collect datasets from YouTube, COCO, and Roboflow with diverse
traffic and accident scenarios.
STEP 2. Data Cleaning- Filter low-quality images, standardize formats, and apply data
augmentation, including:
• Rotation: Randomly rotate images.
• Flipping: Random horizontal or vertical flipping.
• Cropping: Crop and resize images (640x640).
STEP 3. Annotation- Label images using tools like LabelImg with bounding boxes as
required.

16
Figure 3: Annotated images with bounding boxes for accident and non-accident

17
 No. of Images

Total 4022

Train 2815 (70%)

Valid 603 (15%)

Test 604 (15%)

Figure 1a: Data Split Distribution from Online Sources

18
2. Comparative Analysis of Deep Learning Algorithms
STEP 1. Model Selection- Compare YOLO (v5, v8), and CNN for real-time detection tasks.
STEP 2. Training and Testing- Train models on the dataset and evaluate their accuracy and
speed in detecting accidents.
STEP 3. Performance Evaluation- Analyze precision, recall, and the processing time we
experienced to assess model suitability.

19
Figure 4a: Performance metrics of YOLOv8
20
Figure 4b: Confusion matrices of YOLOv8

21
 Metric Start (Epoch 1) Mid (Epoch 5) End (Epoch 10) Key Observation

train/box_loss 1.5815 1.4656 1.089 Steady decrease, showing improved localization.

train/cls_loss 2.5592 1.7172 0.99509 Model gets better at classifying objects over time.

train/dfl_loss 1.6341 1.5761 1.303 Bounding box precision improves consistently.

metrics/precision(B) 0.19477 0.5897 0.80166 Significant increase, fewer false positives.

metrics/recall(B) 0.29492 0.50905 0.74349 Model identifies more relevant objects over epochs.

metrics/mAP50(B) 0.17351 0.53857 0.82517 General detection performance improves significantly.

metrics/mAP50-95(B) 0.08058 0.32341 0.57599 Harder metric, but consistent improvement.

val/box_loss 1.7572 1.5073 1.1337 Decreases, shows improved validation performance.

val/cls_loss 3.0436 1.7918 1.0129 Stabilizes after fluctuations, improving generalization.

val/dfl_loss 1.9647 1.6696 1.3643 Validation bounding box precision increases steadily.

Figure 4c: Training and Validation Losses Across Epochs with Corresponding Precision,
Recall, and mAP Metrics for YOLOv8

22
Figure 5a: Performance metrics of YOLOv5
23
Figure 5b: Confusion matrix of YOLOv5

24
 Metric Start (Epoch 1) Mid (Epoch 5) End (Epoch 10) Key Observation
train/box_loss 0.089858 0.054813 0.04646 Decreasing, indicating improvement in object localization
train/obj_loss 0.038482 0.034282 0.032278 Slowly decreasing, suggesting gradual improvement in object detection
train/cls_loss 0.026188 0.014413 0.0092758 Significantly decreasing, indicating improved classification accuracy
metrics/precision 0.0034738 0.32194 0.54957 Drastically increasing, suggesting improved accuracy of positive predictions
metrics/recall 0.5037 0.32799 0.50413 Relatively stable, indicating consistent ability to detect positive instances
metrics/mAP_0.5 0.0072273 0.25768 0.52306 Significantly increasing, indicating overall improvement in object detection performance
metrics/mAP_0.5:0.95 0.0016676 0.10631 0.25961 Increasing, but at a slower rate than mAP_0.5
val/box_loss 0.075289 0.048714 0.041302 Decreasing, indicating improvement in object localization on validation set
val/obj_loss 0.023829 0.019828 0.017702 Slowly decreasing, suggesting gradual improvement in object detection on validation set
val/cls_loss 0.024045 0.015966 0.0079618 Significantly decreasing, indicating improved classification accuracy on validation set

Figure 5c: Training and Validation Losses Across Epochs with Corresponding Precision,
Recall, and mAP Metrics for YOLOv5

25
Figure 6a: Performance metrics of CNN
26
Figure 6b: Performance Metric and Confusion Matrix of CNN

27
 Average Precision Average Recall

YOLOv5 YOLOv8 CNN YOLOv5 YOLOv8 CNN

0.28545048 0.550393 0.88491 0.342309 0.514946 0.70104

Figure 7: Result of Comparative Analysis

28
TECHNICAL APPROACH

29
Technical Component
Accident Detection YOLO is used for object detection, identifying vehicles and accidents in
with YOLO real-time from CCTV footage. Optimized to detect collision points and
fhfhfhfhfhfh predict accident occurrences with high accuracy.
Tracking Movement YOLO’s bounding box outputs are processed using Optical Flow to track
and Mitigating vehicle movement between frames. This mitigates occlusions and enhances
Occlusions(Optical detection accuracy in complex traffic scenarios.
Flow) jffhfhfhffhhfhfhfhfhfhfhhfhfh
Severity Assessment LSTM models temporal dependencies by analyzing the sequence of events
with LSTM leading up to and following the accident. The recurrent structure of LSTM
fjfjfjfjfjfjfjfj enables better severity classification based on event progression.
Real-Time Dispatch After detecting an accident and assessing its severity, the system prioritizes
Prioritization EMV dispatch in real-time. Ensures ambulances are dispatched to the most
hghghghgh critical cases first, improving response time for severe accidents.

30
EARLY STAGES RESULT

31
Figure 8: Accident Detection

32
Figure 9: Output of Vector Tracking and accident detection

33
Figure 10: Output of accident detection in
YOLO
34
IMPACTS AND BENEFITS

35
 Impact & Benefits
Reduction in Fatalities By prioritizing severe cases, the system reduces response time for critical
hh accidents, helping to lower road accident-related fatalities.
Enhanced Efficiency in Emergency resources are dispatched more effectively, ensuring critical cases
Resource Allocation receive timely attention and reducing delays in emergency response.
Scalability and Global The system is adaptable for various cities and countries, providing a scalable
Impact solution to reduce road accident deaths on a global scale.
Cost Savings for Optimized dispatch reduces operational costs for emergency services and
Healthcare Systems healthcare facilities by improving resource management.
Improvement in Public Faster response times and accurate dispatch prioritization build public trust
Safety in emergency services and improve overall road safety.
Long-Term Societal Fewer accident-related fatalities and injuries contribute to a healthier
Benefits population, lowering long-term healthcare burdens and enhancing
ghhhhhhhhhhh productivity.

36
 REFERENCES
[1] O'Keeffe C, Nicholl J, Turner J, et al Role of ambulance response times in the survival
of patients Emergency Medicine Journal 2011;28:703-706.
[2] Correspondent, Road accidents top cause for youth's, kids' death globally; India loses
1.5 million lives every year: FICCI-EY report, (2023, July 18), Economic Times,
https://economictimes.indiatimes.com/news/india/road-accidents-top-cause-for-youths-kids-
death-globally-india-loses-1-5-million-lives-every-year-ficci-ey-report/articleshow/1018650
11.cms?from=mdr
[3] Diwan, T., Anirudh, G. & Tembhurne, J.V. Object detection using YOLO: challenges,
architectural successors, datasets and applications. Multimed Tools Appl 82, 9243–9275
(2023). https://doi.org/10.1007/s11042-022-13644-y
[4] Ralf, C., Staudemeyer., Eric, Rothstein, Morris. (2019). Understanding LSTM -- a
tutorial into Long Short-Term Memory Recurrent Neural Networks. arXiv: Neural and
Evolutionary Computing, 37
[5] Song, Wang., Zengfu, Wang. (2019). Optical Flow Estimation with Occlusion Detection.
Algorithms, 12(5):92-. doi: 10.3390/A12050092
[6] Ultralytics. Model Structure. 2024. Available online:
https://docs.ultralytics.com/models/
(accessed on 14 October, 2024).
We’ve come to the end of the presentation
THANK YOU

39