Image credit: Medium

Event-driven Pedestrian & Vehicle Detection

Please click on the ‘Video’ and ‘Code’ buttons displayed above to view project source code and model in action!

Abstract:

Most of the deep neural object detection/classification models such as R-CNN, Faster R-CNN & SSD require significantly higher computational power to achieve respectable FPS & Accuracy. But, we need a different cnn-based neural architecture to cater the scenarios where the object detection task is to be executed in real-time rather than waiting for hours to process a 10 min video stream or to run the detector on limited/low-power hardware. Thus, in this project, we implement a real-time event-driven Vehicle Detection & Counting system using the OpenCV’s Yolo v3 (608 x 608) implementation. The model detects and counts pedestrians/vehicles that cross a line of interest projected in any axis within the video stream.

Implemented Tasks:

  1. Experimented & contrasted different YOLO v3 configurations.
  2. Optimized Object Detection & Classification for large video streams.
  3. Implemented vehicle density counter at per-frame/video level.
  4. Developed efficient frame-level per-class object counter.
  5. Designed Event-based object counter to detect incoming & outgoing vehicles in different lanes.
  6. Refactored and optimized project code.
  7. Envisioned an End-to-End system by implementing a real-time object detection progress bar.
  8. Set-up, executed & analysed the model on Google Cloud Platform and verified the overall model accuracies/fps.
  9. Sped-up OpenCV‘s video writer output by utilizing different video extensions.
  10. Fixed FFmpeg dependency issues in Ubuntu Docker container.

Instructions:

Execute in Windows/Linux:

  1. Install Python 3
  2. Install PIP
  3. Install project dependencies: “pip install -r Requirement.txt”
  4. Download Yolo v3 Weights (608x608) from this link. Could not store them in GitHub repo because of file size restrictions!
  5. Save the “yolov3.weights” (downloaded in the above step) file inside the “yolo-coco” folder within the project directory.
  6. Download a sample input video from the link
  7. Save the “overpass.mp4” file inside the videos folder within the project directory.
  8. Run the object detector model inside the project directory: “python Yolo_Detector.py -i videos/overpass.mp4 -o output/final_output.avi

Execute in Docker Container:

Information: The final project docker image is publicly hosted in the Docker Hub portal: https://hub.docker.com/r/poojiyengar5/computer_vision

For Beginner Users:

[Who just wants to run the Object Detector once and see the results - cannot modify or view the code, so don’t use this command if you want to continue the project development in the future!]

  • Command: docker run –rm –name Akshay_Container -ti poojiyengar5/computer_vision:latest python3 HIS.py -i videos/overpass.mp4 -o output/final_output.avi

For Advanced Users:

[For further development or modification of the project code]

  • Please follow the following steps to run the project:
    • Step 1: Container creation

      • Command: docker run –name Akshay_Container -ti poojiyengar5/computer_vision:latest
    • Step 2: Start the container

      • Command: docker start Akshay_Container
    • Step 3: Log-in to the container

      • Command: docker exec -ti Akshay_Container /bin/bash
    • Step 4: Run the Object Detector

      • Command: python3 HIS.py -i videos/overpass.mp4 -o output/test.avi
    • Step 5: Modify the Object Detector Code

      • Command: vi HIS.py
    • Step 6: Exit from the container

      • Command: exit
    • Step 7: Stop the container

      • Command: docker stop Akshay_Container
    • Step 8: Delete the container

      • Command: docker rm Akshay_Container

Future Works:

  1. Post object detection generate detailed vehicle information reports for further statistical analysis.
  2. Exploit pre-trained YOLO models from TensorFlow/PyTorch instead of OpenCV.
  3. Rather than downloading OpenCV binaries from PIP/APT, compile OpenCV source code manually to enable CUDA backend.
Akshay Joshi
M. Sc. Data Science & Artificial Intelligence