Commit
•
96d85bf
1
Parent(s):
ae533b9
Update README.md
Browse files
README.md
CHANGED
@@ -5,7 +5,7 @@ Problem statement and Need for Damaged Roads Detection Model
|
|
5 |
Object modelling for real-world road damage detection in real-time involves a comprehensive workflow. Initially, an image processing model identifies various road damages, including potholes, oil spillage, fallen trees, and emergency events. Simultaneously, the system captures geographical coordinates (latitudes and longitudes) of the detected damages. This data, along with annotated images, is processed to trigger an automated email alert to relevant government agencies. The email serves as immediate feedback, enabling swift actions for road maintenance and hazard mitigation. Additionally, the workflow incorporates anomaly detection for identifying unexpected events such as heavy metal spills or mudslides. By seamlessly integrating image recognition, geospatial analysis, and automated communication, this approach creates a context-aware system capable of addressing road infrastructure challenges promptly and efficiently. The real-time nature of the system enhances public safety and enables timely government responses to ensure road network integrity and minimize potential risks.
|
6 |
|
7 |
Overview of the tasks carried out:
|
8 |
-
1. Collection of data set from RDD2022 [1], this data used and annotated for two objects detection (a) “Damaged Road” (b) Street Light
|
9 |
2. The model used was ssd_mobilenet_v2_320x320_coco17_tpu-8. Total of 3 runs were made first run was annotated with polygon annotation, which was not satisfactory, later two runs were with square and rectangular boxes and hyper parameters changes.
|
10 |
3. The data used about 120 images with 120 annotation pascal voc XML files.
|
11 |
4. Tensor board was captured at the beginning and at the end of the model training where there is no further learning is happening (flattened learning curve).
|
|
|
5 |
Object modelling for real-world road damage detection in real-time involves a comprehensive workflow. Initially, an image processing model identifies various road damages, including potholes, oil spillage, fallen trees, and emergency events. Simultaneously, the system captures geographical coordinates (latitudes and longitudes) of the detected damages. This data, along with annotated images, is processed to trigger an automated email alert to relevant government agencies. The email serves as immediate feedback, enabling swift actions for road maintenance and hazard mitigation. Additionally, the workflow incorporates anomaly detection for identifying unexpected events such as heavy metal spills or mudslides. By seamlessly integrating image recognition, geospatial analysis, and automated communication, this approach creates a context-aware system capable of addressing road infrastructure challenges promptly and efficiently. The real-time nature of the system enhances public safety and enables timely government responses to ensure road network integrity and minimize potential risks.
|
6 |
|
7 |
Overview of the tasks carried out:
|
8 |
+
1. Collection of data set from RDD2022 [1], this data used and annotated for two objects detection (a) “Damaged Road” (b) Street Light.\n
|
9 |
2. The model used was ssd_mobilenet_v2_320x320_coco17_tpu-8. Total of 3 runs were made first run was annotated with polygon annotation, which was not satisfactory, later two runs were with square and rectangular boxes and hyper parameters changes.
|
10 |
3. The data used about 120 images with 120 annotation pascal voc XML files.
|
11 |
4. Tensor board was captured at the beginning and at the end of the model training where there is no further learning is happening (flattened learning curve).
|