Reaching out the mentors of Project 16 – Traffic Intersection Monitoring System

[BhaskarJoshi-01]

Dear @openvino-dev-samples, Dan Qiu and Yiwei Lee,
I’m Bhaskar, and I have hands-on experience in computer vision, object detection, deep learning, and real-time optimization.

Key Highlights of My Background:
Academic Qualifications: B.Tech (Hons) in Computer Science + MS by Research in AI GPA: 9.27/10 (Top 2%) Dean’s List for six consecutive semesters
Research Interests: Deep Learning, Reinforcement Learning, LLMs, and Generative AI
Publications and Patents: 5 IEEE research papers published, 1 journal paper under review 3 US patents (1 accepted, 2 under review) Onsite Research in my undergrad and published a research paper with York University and Université Paris Dauphine-PSL (completed in 3-4 months)
Industry Experience: Adobe, Airbus, Intel (worked on pti-gpu) , and Google Summer of Code at Internet Archive Teaching
Research Assistant Roles: Teaching Assistant in Machine Learning, Reinforcement Learning, Operating Systems, Networking
Research Assistant at Robotics Research Center, IIIT Hyderabad in Deep RL

I’m very interested in Project 16 – Traffic Intersection Monitoring System with OpenVINO and would love to contribute to its development.
As I prepare my proposal, I’d like to clarify a few key details to ensure I’m aligned with the project’s objectives:
Object Detection Models:
Are there any preferred object detection models for this project (YOLO, Faster R-CNN, SSD), or would experimentation with multiple architectures be beneficial?
Would we focus on pre-trained models, or is there room for training custom models?

Violation Detection:
What types of traffic violations are a priority (e.g., red-light running, illegal lane changes, speed estimation)?
Should we incorporate multi-camera setups for enhanced tracking and cross-intersection monitoring?

Performance Optimization with OpenVINO:
Are there specific OpenVINO optimizations or quantization techniques you recommend for real-time performance?
Should the focus be on edge devices (e.g., Intel NCS, Movidius) or general-purpose CPUs/GPUs?

Dataset & Evaluation Metrics:
Would this project use public datasets (e.g., Cityscapes, Berkeley DeepDrive) or require custom data collection?
What evaluation metrics should be prioritized for accuracy and efficiency (e.g., mAP, inference latency, FPS, false positive rate)?

Graphical User Interface (GUI) Development:
Should the GUI support live video feeds and interactive analytics, or is a static reporting dashboard sufficient? Are there specific Qt design guidelines or components to follow?
Additionally, would developing an initial prototype be useful for demonstrating feasibility in the proposal? I’d really appreciate any guidance you can provide, and I look forward to discussing how I can contribute effectively to this project.
I believe my background and skills make me a strong candidate for this project, and I am committed to delivering high-quality results.

Best regards,
Bhaskar
https://www.linkedin.com/in/bhaskar-joshi-968a591a4/

[BhaskarJoshi-01]

@dan20210809 can you kindly provide suggestions? I am super excited to contribute to the project and maybe the post skipped your attention.

[dan20210809]

Hi, Bhaskar, thank you for your questions, please see my comments below:

Object Detection Models:
We would choose pre-trained model in this project, YOLO would be fine. Re-training custom model is not necessary, because we want to quantize and analyze the model performance in the pipeline, custom model may have issues when quantizing.
Violation Detection:
This project is the vision part of V2X(Vehicle to Everything) solution, in which 4 cameras and 4 radars will be used to determine the violations. We are going to use 4 cameras and ignore the adverse situations (like heavy rain) to simplify the pipeline so that we can focus on the deployment skills and showcase how to increase the inference efficiency of the project. In this case, we want to at least detect red-light running and speeding.
Performance Optimization with OpenVINO:
Our demo platform will be on ultra125h which include CPU, iGPU and NPU, and we are going to deploy the pipeline across CPU, iGPU and NPU to find the best solution. Movidius is acceptable, but it may not enough to run this pipeline. We are going to use async API of OpenVINO and NNCF to quantize the model.
Dataset & Evaluation Metrics:
We are not going to re-train the model, and the metrics will be the FPS and mAP ( mAP would not be a very hard requirement cause we don't have complete data in the V2X solution which would impact the accuracy).
Graphical User Interface (GUI) Development:
The GUI is expected to support live video feeds and interactive analytics. Qt-python is a widely-used GUI components, you can find many study contents in the web. We may have a window to real-time display the results, and an alert for violation, also we want to show statics of a required time period, for example we can check how many violation times and which is the top type.

The technology used in the project is very mature, if new idea(for example LLMs) can be used will be a plus. Looking forward to your proposal, feel free to contact me.