Skip to content

Latest commit

 

History

History
92 lines (59 loc) · 3.77 KB

README.md

File metadata and controls

92 lines (59 loc) · 3.77 KB
Raw

Video Surveillance for Road Traffic Monitoring

The goal of the project is to provide a robust system for road traffic monitoring using computer vision techniques for video analysis.

Team Members

Roger Casals María Gil Oscar Mañas Laura Mora
[email protected] [email protected] [email protected]

Execution

Execute the program as follows:

python main.py -h
usage: main.py [-h] [-w WEEK] [-t TASK]

M6 - Video Analysis: Video Surveillance for Road Traffic Monitoring

optional arguments:
 -h, --help            show this help message and exit
 -w WEEK, --week WEEK  week to execute. Options are [1,2,3,4,5]
 -t TASK, --task TASK  task to execute. Options depend on each week.

Figures for the results will be saved in results/

Week 1

The first step of the project is to implement the evaluation metrics needed to prove the robustness of the tool. We implement metrics for:

  • Object Detection:

    • Mean Intersection over Union
    • Mean Average Precision

  • Optical Flow:

    • Mean Square Error in Non-occluded areas
    • Percentage of Erroneous Pixels in Non-occluded areas

Slides

Week 2

The goal of this week is to model the background of a video sequence in order to estimate the foreground objects. The background is modelled using a single gaussian method, both non-adaptive and adaptive.

For this, we use grayscale videos as well as colour information from different color spaces. We also compare our method with several SOTA methods from public python libraries.

Slides

Week 3

The goal of this week is to perform object detection for car and from those detections track each one of the cars independently.

  • Object Detection:

    • Mask-RCNN and Faster-RCNN are used "off-the-shelf", as they are trained with the COCO dataset (which provides with the necessary class car)
    • Fine tuned Mask-RCNN to our dataset.

  • Tracking:

    • By overlap, we consider the IoU between boxes in consecutives frames in order to assign a unique id to a car
    • Using the Kalman filter

Slides

Week 4

The goal of this week is to estimate the optical flow of a video sequence and use it to stabilize any potential camera jitter and improve an object tracking algorithm.

  • Optical Flow with Block Matching
  • Video Stabilization with Block Matching
  • Object Tracking with Optical Flow

Slides

Week 5

During week5 we aim to solve Track 03 of the AICity Challenge 2020: City-Scale Multi-Camera Vehicle Tracking. The goal is to track vehicles across multiple cameras both at a single intersection and across multiple intersections spread out across a city. This helps traffic engineers understand journey times along entire corridors.

  • Multi-target single-camera (MTSC) tracking

    • Tracking from detections in Yolo3
    • Discard small bounding boxes
    • Filter static and momentary tracks

  • Multi-target multi-camera (MTMC) tracking

    • Global track representations extracted with Metric Learning
    • Track comparison reduction by considering spatio-temporal coherence
    • Vehicle re-identification: exhaustive/consecutive/graph based

Slides

Report