Smart Intersection - daily traffic flow

Project Website

​https://bzz3ru.wixsite.com/smartintersection

Project Objective

The goal of this project is to create a method for estimating the statistics for vehicle count/traffic flow into one intersection in New York City. As an example, record videos of the northbound traffic on Amsterdam Avenue, as vehicles are entering the 120th St./Amsterdam Av. intersection. Using YOLOv3 deep learning model, detect and count vehicles as they approach/enter the intersection from south, making sure that there is no double-counting. Use 180 second long video fragments (approximately two traffic light cycles), and repeat up to half a dozen times a day, for a number of workweek/weekend days during the same times of each day. Compare the vehicle count (traffic flow) as a function of the time of the day. Utilize NVIDIA DeepStream deployed on COSMOS GPU compute servers to run the model. The method should be generalizable/expandable to any direction of vehicle movement, when appropriate camera views are available.

Reading Material

Week 1 Activities

• Get ORBIT/COSMOS account and familiarize oneself with the testbed procedures
• Learn about YOLOv3 deep learning models for object detection
• Read about NVIDIA DeepStream
• Explore the image (set of computing tools) available on COSMOS, which uses DeepStream and can deploy YOLOv3
• Record and save 6 videos during one day (to be repeated when the method is debugged and fully functional)
• Brainstorm about vehicle counting/traffic flow estimation methodology

---

Week 1 Weekly Meeting Presentation: ​https://docs.google.com/presentation/d/1Sf9hzpo3WQsEPwbhKfic2xWCskH1EViD-3SNb_foouA/edit?usp=sharing

Week 2 Activities

• Understand the concepts of object detection in 3D Point Cloud
• Gain an understanding of NVIDIA’s DeepStream SDK
• Get comfortable deploying YOLOv3 on the COSMOS testbed
• Use existing datasets to play around with DeepStream and YOLOv3

---

Week 2 Weekly Meeting Presentation: ​https://docs.google.com/presentation/d/1Cl8MbsSU3ZAq5lpRuE0eVBSwnIX4jTUAci5uAgP7Vt8/edit?usp=sharing

Week 2 Team Meeting Presentation: ​https://docs.google.com/presentation/d/1O2yCze4fmVOAFGCi0u6WTZq8VTFhLc_J448skeygguw/edit?usp=sharing

Week 3 Activities

• Investigate existing RGB-D (RGB + depth map) object detectors whose models we can immediately put to use for inference
• Look into existing 3D Point Cloud object detection implementations
• Learn how to run DeepStream's YOLOv3 implementation
• Investigate DeepStream Python bindings for use with YOLO

---

Week 3 Weekly Meeting Presentation: ​https://docs.google.com/presentation/d/13vqiw0kkyT0_XPzPv3NiowIvxc22SapfxM_ZbAKn1Cc/edit?usp=sharing

Week 3 Team Meeting Presentation: ​https://docs.google.com/presentation/d/1jwq6h05mw1vHt6_C1Br4MM_0LQDZRlIEoSvEasJ5Rsg/edit?usp=sharing

Week 4 Activities

• Investigate YOLOv4 and its use with TensorRT
• Look into getting output/data processing based on the outputs from DeepStream
• Look into the DeepStream tracker to build on top of
• Build a presentation slide set to inform the intern class about DeepStream and YOLOv3

Introduction to DeepStream and YOLOv3 Presentation Slides: To be posted at a later date

---

Week 4 Weekly Meeting Presentation: To be posted at a later date

Week 4 Team Meeting Presentation: To be posted at a later date