Sigmoidal Success Story
pre-processed brief videos
What was the business problem?
Today, most criminal activities are carried out using handheld weapons. Many studies have revealed that handheld weapons are the most important criminal elements used for various crimes, such as theft, illegal hunting, and terrorism. The solution to reduce such criminal activities is installing a surveillance system or control cameras so that security units can take appropriate measures at the early stages.
Weapon detection is challenging due to the various subtleties associated with it. The most important problems in weapon detection are self-occlusion and the similarities between objects and background structures. Self-occlusion occurs when a part of the gun is blocked on one side. The similarity between objects occurs when different objects such as hands and clothes look like weapons. Background problems refer to those related to the background against which the gun is located.
Our client needed a software solution that detects humans on a video stream in real-life which will be as lightweight as possible to be run on a low-cost, small device equipped with an Nvidia GPU.
What savings and profits does the client achieve?
We establish a proactive system, that enabled our client to pre-emptively alert the security immediately after detecting threatening objects, prompt action could be taken to stop the potential criminal from committing a crime. Our client was able to automatically analyze multimedia events and calculate security levels.
Due to the usage of VGGNet architecture, our client could detect and classify weapons in various images and real-time videos. Our solution reduced potential criminal activities and allowed for much quicker detection and reaction.
How did we accomplish it?
Deep learning is a sub-field of machine learning. It uses many layers of non-linear processing units for deep learning and feature extraction and conversion. The deep learning structure is based on the learning of more than one feature level of data.
The hardware capabilities of Nvidia Jetson made challenging to utilize Deep Learning’s full range of abilities. We managed to provide a performance-optimized, OpenCV-based algorithm that worked in real-time, with very low latency and with the required precision. To speed up computations, we’ve used the C++ OpenCV library in our Python code. The official Python version didn’t provide access to CUDA algorithms, so we created a dedicated library to provide appropriate bindings.
Sigmoidals’ engine was engineered in 5 general stages:
- Human Detection
- Optimization for low light conditions
- Activity Detection
- Deployed on Nvidia Jetson
- Implemented in security cameras
During this project Sigmoidals’ team, engineered many extensions to the existing algorithm to enable more comprehensive predictions like real-time body part segmentation and gesture recognition.
Our solution based on Deep learning and computer vision allowed us to:
- Draw contour over a human body
- Leverage bloom filters and advanced mathematical models
- Process real-time video
- Handle difficult lighting conditions
In addition, our weapon classification system based on the VGGNet architecture was used to detect and classify weapons in the various images, allowing us to automatically analyze multimedia events and calculate security levels.
How did we boost the project with ?
Sigmoidal utilized our Data-Centric approach to maximize the effectiveness of the software. We systematically altered and improved the datasets in order to increase the accuracy of our software.
Sigmoidal utilized the Blue-Green deployment method allowing us to release a model transferring production traffic from the old environment to a new one, both of which were running in production.
We performed metrics definition and monitoring to assess our model performance every step of the way. Our method consists of the calculating distance between predicted and baseline data.