Vision Semantics Ltd. 

VSL Inside: AI that Finds People in Time

Artificial Intelligence Products built on World Leading Research
Deep learning neural networks which transform the outcomes of vision analysis  

Nearly two decades ago, we set out to create products that would transform the way computer vision analytics was applied to real world problems. Today, our patented ‘VSL Inside’ suite of products enable person re-identification (Re-ID) to be deployed for the most critical Smart City and Public Safety applications across the world to solve problems and save lives.

The products are based on world leading research led by Professor Sean Gong at Queen Mary University of London and co-founder of Vision Semantics. The Queen Mary Computer Vision Group has been conducting world-leading research in computer vision and machine learning for over 20 years, and is internationally renowned for its work on unusual behaviour recognition, action recognition, person re-identification, multi-camera tracking, and face analysis in video and images. More latterly recognized as the leading research group active in domain transfer learning and unsupervised learning for Re-ID. This research has led to over 300 published paper, over 23,000 citations and filing of 6 families of international patents.

 

Vision Semantics have applied the best engineering practice to scale and operationalise our solution over the last 10 years. The systems have been commercialized and tested in live operational situations with government and commercial agencies including the Department of Trade & Industry, Ministry of Defence, Defence Science and Technology Laboratory (DSTL), the Home Office, and industrial partners including BAE Systems, BAA, BT Labs, BBC R&D, and QinetiQ.

AI Powered Person Re-Identification

The software platform is designed to significantly enhance video analysis in public safety and smart city applications. The video surveillance industry is experiencing a massive growth spurt as we increasingly use cameras to prevent crime, protect our property, and improve public safety.  However, data generated from digital surveillance cameras has become so massive that it is impossible for human operators to make sense out of it in a timely manner.

A major challenge is Person Re-Identification (Re-ID). That is the process to find a matching person (in a gallery of images) for a given image using not (just) facial features (which are often obscured) but features found on the entire body (like clothing, height, carried objects, etc).  Face recognition works only when the subject is close enough, well-lit and facing towards the camera. Usually in CCTV footage this is not the case, and an individual can deliberately cover their face from the camera.

Vision Semantics provides a unique Re-ID solution including human-in-the-loop relevance feedback (supported learning), active learning and video summarisation with attributes search. As a result, when video imagery is low-resolution and face recognition is no longer feasible (less than 10% surveillance videos contain near frontal view facial imagery data), then Vision Semantics products provide the solution. This makes the Vision Semantics product unique due to its scalability and ability to search for people when faces are not visible in unstructured and uncontrolled environments (most). Vision Semantics can analyse big video data very rapidly (100x faster than real-time) and re-identify people more accurately than highly trained human experts, especially when the video data is from unknown sources without domain knowledge for search optimisation.

VSL Inside:  AI Person Re-Identification

Vision Semantics through an implementation ‘VSL inside’ provides a best of breed person Re-Identification (Re-ID) module based on superior algorithms and techniques created by the leaders in the academic field. Prof Gong is a world-leading authority in computer vision for person re-identification and visual attributes search in large scale video data. His research has developed patented technologies that underpin Vision Semantics software for:

  • Large scale person re-identification by relative distance ranking

  • Video-tracklet Re-ID matching

  • Transfer and unsupervised learning Re-ID

  • Human-in-the-loop relevance feedback learning and active learning for Re-ID

  • Imbalanced attributes learning Re-ID

  • Open-set Re-ID: Person re-identification over large scale unknown target domains in open-world search beyond closed-set benchmark training and testing data

  • Fast & efficient semantic summarisation and search of videos by attributes

 

Vision Semantics have implemented a patent protected Dynamic Deep Reinforcement Active Learning algorithm within a Convolutional Neural Networks which provides the world class performance inherent in Vision Semantics Re-ID:

  • World leading person re-identification research and commercial grade software

  • 20-fold faster than real-time in video search

  • 1,000-fold faster than human experts. Detect & Re-ID at 334 fps (3ms/frame)

  • 2.4-fold more accurate than human experts (Top-20 Re-ID precision 98.6%)

 

Accuracy, Scalability, Speed

Accuracy

  • 2.4-fold more accurate than human experts (Top-20 Re-ID precision 98.6%)

  • Rank-1 Re-ID rate 85.71% in unknown target domain (without training data) vs. SOTA 63.7% (unsupervised learning Re-ID)

Scalability

  • Configurable distributed processing (multi-GPUs, multi-threaded server implementation)

  • Commercial system with customers (TRL9 maturity)

Speed

  • 20-fold faster than real-time in video search

  • 1,000-fold faster than human experts, detect & Re-ID at 334 fps (3ms/frame)

© 2019 VISION SEMANTICS LTD