COVID-19 Curfew and Social Distancing Enforcement using AI enabled Drones
Amarjot Singh, PhD
Founder & CEO
SkyLark Labs LLC

The SARS-CoV-2(also known as COVID-19 or Novel Corona Virus) infectious outbreak has rapidly spread across the globe and precipitated varying policies to effectuate physical distancing to ameliorate its impact.

With over 2.28 million active cases all over the world and over 36,000 cases in India alone, it’s a tough job for the Police to inspect each and every corner of the street for people violating the social distancing policy and lockdown implemented by the Government of India.

Skylark Labs is assisting the Police in Punjab, India with their AI enabled Drones to enforce the curfew and maintain social distancing.

Skylark Labs using drones to enforce social distancing and curfew in Punjab, India.

Here are some of the main features of this system, which stand out from other approaches in solving this task:

Human detection:

  • It’s no surprise that Computer Vision using Deep Supervised Learning algorithms is performing better than humans(since 2013) in tasks like object recognition, detection, reading obfuscated text, others, given enough data and compute power.
Humans being detected using Drones, and categorizing them based on their proximity to other humans
  • The System uses a Deep Neural Network which is fully convolutional (YOLO based), trained on annotated data consisting video sequences from drones. Using a fully convolutional network helps the system run as varied image sizes.
  • Training the system is a traditional Object Detection task. For this particular scenario, Object detection turns out to be: given an image locating the position of the humans present at various locations and placing a bounding box around them.
  • Each Image(frame from the video) is divided into a grid(each image having multiple grid cells). The system uses Anchor boxes(also known as Prior boxes) to detect multiple humans in each grid cell.
Anchor boxes

Each anchor box carries its own particular x, y(centre coordinates) and h, w(height and width).

Instead of predicting arbitrary bounding boxes in each grid cell, the system predicts offsets relative to each anchor box initialized. Use of Anchor boxes eliminates the need to scan an image with a sliding window that computes a separate prediction at every potential position. This helps detection of multiple humans in single grid cell.

  • The system uses IOU (Intersection over Union) to predict how likely there is a human in the corresponding box given the ground truth.
  • The Anchor boxes are selected based on high confidence scores and the rest are eliminated using NMS(Non-Maximum Suppression).
  • Whole system has been trained end-to-end. As the system doesn’t use a region proposal based method(bit on the computationally expensive side)for detecting humans, it can run real-time.
  • At test time, each frame of the real-time video passes through the pre-trained Convolutional Network and gives the respective accurate bounding box for each detected human.
Human detection using drones during night
  • As the system was trained on wide range of videos(being shot during day and night), it can perform Human detection even in low-light, showing its robustness.

Invariant towards Long Range detection:

Detecting people from long range and sensitive towards proximity of detected humans
  • The system detects humans even from a long range(upto 1000ft.), which is helpful to detect large number of people in a single frame. This decreases monitoring of drones constantly by zoning in for humans being detected.
  • The system also performs accurately in detecting humans based on their closeness to other humans. For instance, as shown in the above visual, a person is in green box as he’s not being surrounded by any other human nearby, whereas a group of 2 on motorbikes are in red boxes showing them being close to each other, placing them in unacceptable category(distance<6ft).

Real-Time Detection:

Command Centre for visualization of all the drones capturing the videos and detection being done at the same time. The system also gathers statistics on number of people, and the alerts.

The proposed system performs the computation and memory demanding distance proximity processes on the cloud, while keeping short-term navigation onboard.

This image depicts the usage of the system in real-time. On the left, we have standard drone video sequence with people being detecting(around the bounding boxes). And on the right, we have plots showing number of people detected at various time intervals and alerts being sent to the Police.
  • This helps the system to perform tasks in real time.
  • To back this statement, the drone captures the video at 25 fps and sends it for processing the video sequences. The system does its computation, bounding box detection at rates of nearly 45 fps(close to 22ms for each image).

Real-time Alerts:

Police officers use these alerts sent to their mobile phones about the accurate location of the suspect.
  • Once a human has been detected violating the curfew, SMS alerts are sent right away to mobile phones of the Police officers present closest to the spot regrading the location of the suspect and the its corresponding real-time video.
  • The location gets captured by the drone itself.


  • Skylark Labs is using robust Deep Learning algorithms to solve real-world problems, with high accuracy and real-time processing which is really helpful for current state of pandemic.
  • The Police can get greatly benefited from this system constantly monitoring people, and at the same time can stay safe from exposure to this deadly virus.

Skylark Labs develops continuously evolving and customizable threat detection systems for enhanced physical security.

Recent Posts

View All Posts