Mentors : Prof. Mazad Zaveri, Prof. Mehul S Raval
Mentors : Prof. Mazad Zaveri, Prof. Mehul S Raval
DAC Members: Prof. Jayendra M Bhalodiya, Prof. Kuntal Patel, Prof. Rajesh S Gujar
RPS 0 (research proposal presented to the DAC committee)
Research Questions
Can we use other (vision-based) modality to overcome the limitations of the CCTV-based system? If so, then
Will it help to study “Others” violations?
Will studying the link between human error and road conditions help?
Will it be possible to assess existing road conditions through the iRAP methodology?
Is it deployable on resource-constrained devices?
Will it be a real-time solution?
Expected Outcomes
Aerial dataset of vehicles (Indian road scenario) for detection and tracking.
Traffic rule violations detection model.
Dictionary of MVDR/MVA rules.
Method to investigate the linkage between human error and road condition.
Framework to quantify driving behaviour.
Method for traffic flow analysis.
Aerial dataset of road attributes (iRAP methodology).
Framework for road survey (iRAP methodology).
Comparative study on models' performance (workstation vs. resource-constrained devices).
RPS 1 (research progress seminar presented to the DAC committee)
Problem statement
Vision-based investigation of road traffic violations in India using UAV video.
Solution
Research work on road traffic investigation at an Indian urban roundabout is published in the Transportation Engineering journal titled "Vision-based investigation of road traffic and violations at urban roundabout in India using UAV video: A case study." https://doi.org/10.1016/j.treng.2023.100207
Acknowledgement/Award
RPS 2
Problem statement
A Cyber-Physical System for Road Traffic Monitoring using UAVs
Solution
We proposed a “UAV-based Urban Traffic Monitoring (U-UTM)” cyber-physical system to automatically detect traffic violations based on vehicle movement and traffic flow parameters, thereby improving road traffic safety.
This research work is presented at the 18th IEEE International Conference on Vehicular Electronics and Safety (ICVES 2024). "U-UTM: A Cyber-Physical System for Road Traffic Monitoring Using UAVs," 2024 IEEE International Conference on Vehicular Electronics and Safety (ICVES), Ahmedabad, India, 2024, pp. 1-6, doi: 10.1109/ICVES61986.2024.10927911
RPS 3
Problem statement
Are Indian drivers follwoing the defensive driving practicles?
Solution
We proposed novel surrogate safety measures (SSMs) and a star rating scheme to evaluate defensive driving behaviour of road users.
Research work related to evaluating defensive driving behaviour is published in the Multimodal Transportation journal titled Evaluating defensive driving behaviour based on safe distance between vehicles: A case study using computer vision on UAV videos at urban roundabout." https://doi.org/10.1016/j.multra.2025.100227
Vehicle detection on Indian Road using HOG, PCA, SVM and Sliding window approach
This project was developed as a part of "Artificial Intelligence" course
HOG was used to extract features of the vehicle. We adjusted parameters of HOG window such that length of feature will be 3780 for any vehicle in the dataset.
We used Principal Component Analysis (PCA) to find out only relevant features out of 3780. We found that only 1500 features were useful.
We trained a SVM classifier(C=10 and kernel= RBF) on such principal features.
Further, to propose regions, we used a sliding window approach. We used 3-level image pyramid and different sizes of windows.
We got maximum accuracy (of 48%) for the "car" class.
We have used the India Driving Dataset (IDD).
We concluded that our region proposal method is not efficient.
Vehicle detection on Indian Road using HOG, PCA, SVM and Sliding window approach (...extension)
This project was an extension of the previous AI project.
Here, we tried different methods of up-scaling and down-scaling for image patches to fit into HOG window size.
With this approach we achieved the F1 score of 94% but still detection accuracy was very low.
Still need to work on the region proposal method.
UAV camera (facing downward) calibration
This was a part of Collaborative Research Project-I. We are using the DJI MAVIC 2 PRO drone.
We used the FOV angle formula to find out Ground Sample Distance (GSD). GSD is the value of actual distance represented by a single pixel on an image.
Accuracy of GSD based calculation is dependent on the bounding box (provided by vehicle detection model).
Another solution, to establish this distance to pixel mapping is create a dataset, of actual vehicle size and its image footprint and then, train a regression model to estimate relationship.
Vehicle detection on Indian Road using aerial imagery
This project was an extension of previous AI projects but with an aerial dataset (VisDrone 2019 detection) instead of IDD, as part of Machine Learning and Computer Vision course.
Here, we used AlexNet to extract features which were previously done using HOG.
This features then forward to the Global Average Pooling (GAP) layer.
With AlexNet classifier, we achieved 80% of test accuracy (tested on VisDrone dataset).
Further for detection, we used selective search to propose regions but this takes 1-2 seconds to do so.
Instead of using separate classifiers and region proposal techniques, we tried the YOLOv5 object detection model for vehicle detection. We trained YOLOv5 on the VisDrone dataset and tested it on our collected traffic video.
We observed misclassification among several classes such as car and van, bus and truck. Still, YOLOv5 is a good solution for such vehicle detection tasks.
Vehicle tracking on Indian Road using aerial imagery
After successful implementation of the detection model (YOLOv5), next we targeted vehicle tracking to complete our pipeline. For testing, we have used Indian road traffic video.
We tried the SORT algorithm for tracking.
We observed multiple tracks of a single vehicle (specifically for two-wheeler and pedestrian). We looked for another version of SORT.
DeepSORT requires a separate dataset known as Vehicle Re-identification dataset for training. We find only one such dataset- VeRI-776 dataset (but it has modality other than aerial). but it was not useful for our case (aerial imagery).
At the end, we concluded that the problem of broken tracks is not only because of SORT. SORT follows the detector so, if the detector misses a vehicle in "some number " of consecutive frames (beyond the SORT parameter value) then SORT will certainly miss that vehicle and will create a new track for the same vehicle. So, choose a better object detection model.
Vehicle detection on Indian Road using aerial imagery with YOLOv7
This work was done to use a better vehicle object detection model and complete the tracking task. (This was a part of my Comprehensive exam)
We explored aerial datasets (AU-AIR and VAID)other than VisDrone 2019 detection dataset but we found that VisDrone dataset has more samples of different vehicle classes.
We trained YOLO on VisDrone and achieved 60% of accuracy. We used YOLOv7-w6 which supports 1280 X 1280 input image.
We again tested a complete pipeline (YOLOv7 and SORT) on our collected traffic video.
We observed that YOLOv7 is faster than YOLOv5 but still facing similar issues (miss-classification and missing small objects in several consecutive frames). One of the possible solutions could be to create our own dataset for Indian road traffic and train the model , again on it.
Video Annotation tool - A LEPC
This project was developed as a part of the "Human Computer Interface" course.
We explored publicly available tools such as CVAT, LabelImg, and VIA for video annotation. All are semi-automatic tools.
We thought of optimizing time to create a dataset with less human intervention. So, we developed our own tool. We also defined different categories of vehicles.
Apart from bounding box and class information, it allows embedding other details in ground truth. It generates ground truth in CSV file format.
This tool will help to create our own dataset efficiently.
Handling shortcomings of Vehicle detection model
Will be updated very soon...