The problem CrowdDistance solves

Post the lockdown period against the coronavirus, and as the country gets back on its feet, it is crucial that the community takes precautionary measures to curb the further spread of the virus. Drastic steps must be taken by the community and organizations alike to prevent a second-wave of COVID cases. Avoiding crowded areas and practicing social distancing must be undertaken by the common citizen and businesses must enforce these norms within their organization’s premises.

We present to you, CrowdDistance, a Deep Learning and GeoLocation based service that enables users and organizations to maintain the appropriate physical distance and avoid contracting the Coronavirus. Our ComputerVision-driven solution caters to both the community and business industries in providing the appropriate analytical tools that benefit each of them.

The solution ingests data from surveillance camera feeds to chart crowd density data and place circular markers on a mapping service that depicts the number of people in an area. Users can take advantage of the crowd data and heatmap visualizations on the map to avoid crowded areas. This will help the community to better plan and avoid highly-dense areas prior to actually traveling to their destination.

Additionally, CrowdDistance provides a Crowd Analytics Dashboard to accurately depict specific social distance violations and alert the security personnel to disperse any emerging crowd. Moreover, organizations can take advantage of Real-time Graphs and Motion HeatMap Visualizations to gather actionable insights in providing a safe environment for their employees. This is yet another step that organizations can take to “Accelerate the New normal” for the post-lockdown days.

Challenges we ran into

The biggest hurdle we faced was in the Core Accelerator of our solution, the Deep Learning Model. We required a powerful Object Detection model in order to detect all the pedestrians in a video feed, even in highly crowded environments. Using pre-trained Tensorflow Object detection models did not make the cut. Since the success of the entire solution depends on how accurate the ML model performs, this stage represented a critical bottleneck in our solution.

To distinguish ourselves from typical github projects on social distancing, we had to retrain the Object Detection model on a dataset tuned for detecting people in highly-crowded environments. We utilized the PyTorch framework for its superiority in model development when compared to TensorFlow. We had to spend a substantial amount of time preparing the dataset for detection in crowded environments.
After countless GPU-training hours and adding several image augmentations into the training pipeline, we managed to achieve highly accurate detections. The Deep Learning model had attained a sufficient precision so as to detect heavily occluded people, babies in their prams and even mannequins inside shopping stores!

Overcoming this major hurdle is what helped us strengthen the robustness of our solution and ensure a swift process workflow for the rest of the solution.