Embedded Files
A Computer Vision project
Datafest Fall 2021
University of Southern California
Team
Dr. Seon Ho Kim
(Associate Director,
Integrated Media Systems Center)
Harsh Jalan
Utkarsh Baranwal
Wonjun Lee
Carlos Zamora
Vibhav Chitalia
Motivation Behind the Project
Motivation Behind the Project
Homelessness has been and is a growing issue in LA County
Overall homeless population (in and out of homelessness) is 66,000 in January 2020.
Tackling homelessness begins with identifying homeless populations and encampments.
Currently, identifying homeless populations and encampments is done through manual counting by LADWP volunteers in Los Angeles.
We are solving this problem by performing machine learning on images to automatically detect homeless encampments and classify them.
Cameras will be set on the vehicles and while moving these vehicles will take images from LA streets. These images will be given to our homeless encampment detection model. The model will detect the number of encampments in images. GPS will be used to detect the location and count the number of encampments in a certain location.
Dataset Details
Dataset Details
857 images of homeless encampments from LA Sanitation Department.
300 images collected from internet.
Adjacent image shows the histogram of dataset against the ratio of area of encampment to the total image area.
Training Data
Training Data
The training data consisted of 80% of total dataset.
Each mosaic is made up of multiple images from training data.
Various augmentation techniques such as HSV augmentation, left-right flipping and image translation are applied to the data during training.
Model Used - YOLOv5 Small
Model Used - YOLOv5 Small
YOLO strands for You Only Look Once.
This algorithm was published in a research article by Redmond et. al in IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR).
YOLO suggested the use of an end-to-end neutral network that makes predictions of bounding boxes and class probabilities all at once.
YOLO divides each image into grids and each grid detects objects within itself.
YOLO performs better in terms of speed and accuracy in comparison to the Faster RCNN and EfficientDet algorithm.
YOLOv5 is the latest version developed by Ultralytics in 2020. YOLOv5 has four models small, medium, large and extra large with different accuracy and speed.
For our purpose, we need to detect images in real time. So, speed plays a significant role. Hence, we have built a YOLOv5 small model for our task.
Experiments
Experiments
Our Team trained multiple YOLO Models with given specifications:
Divided images into three groups:
Images having bounding box area < 3%
Images having bounding box are between 3% and 6
Images having bounding box area > 6%
Used input Image Resolution Size: 200X200, 400x400, 640x640, 700x700, 800x800
Batch size: 32, 64
Optimizer: SGD
Results and Key Inferences
Results and Key Inferences
We can detect tents in the image for most cases using Vision-based techniques (there is still scope of improvements for differently appearing tents).
Aspect ratio for input images for the model should be more closer to square than rectangle.
Recall increases on increasing the image resolution, but after some time, it starts to decrease on increasing the image resolution. Implying we can work with smaller resolution as well.
Page updated
Report abuse