# US 2021/0334542 A1 · Explainable multi-entity event recognition

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0334542 A1
- Document ID: US20210334542A1
- Status: Published application
- Kind: A1
- Application number: 16/861,209
- Filed: 2020-04-28
- Published: 2021-10-28
- Earliest claimed: 2020-04-28
- Assignees: 3D Industries Limited
- Applicants: 3D Industries Limited
- Inventors: Sukrit SHANKAR; Seena REJAL
- Web page: https://patentgrep.com/patent/US20210334542A1

## CPC

- G06V 20/52: Surveillance or monitoring of activities, e.g. for recognising suspicious objects (recognising microscopic objects G06V20/69)
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F 18/2413: based on distances to training or reference patterns
- G06F 18/24147: {Distances to closest patterns, e.g. nearest neighbour classification}
- G06F 18/29: {Graphical models, e.g. Bayesian networks}
- G06N 3/006: based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/092: Reinforcement learning
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06V 20/41: {Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items (segmenting video sequences G06V20/49)}
- G06V 20/46: {Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames}
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

An image processing system has a memory storing a video depicting a multi-entity event, a trained reinforcement teaming policy and a plurality of domain specific language functions. A graph formation module computes a representation of the video as a graph of nodes connected by edges. A trained machine learning system recognizes entities depicted in the video and recognizes attributes of the entities. Labels are added to the nodes of the graph according to the recognized entities and attributes. The trained machine learning system computes a predicted multi-entity event depicted in the video. For individual ones of the edges of the graph, select a domain specific language function from the plurality of domain specific language functions and assign it to the edge, the selection being made at least according to the reinforcement learning policy. An explanation is formed from the domain specific language functions.

## Claims

### Claim 1

An image processing system for recognizing multi-entity activities, the image processing system comprising:
a memory storing a video depicting a multi-entity event, a trained reinforcement learning policy and a plurality of domain specific language functions;
a graph formation module configured to represent the video as a graph of nodes connected by edges;
a trained machine learning system configured to recognize entities depicted in the video and recognize attributes of the entities, such that labels are added to the nodes of the graph according to the recognized entities and attributes, the trained machine learning system also configured to compute a predicted multi-entity event depicted in the video;
a processor configured, for individual ones of the edges of the graph, to select a domain specific language function from the plurality of domain specific language functions and assign it to the edge, the selection being made according to the reinforcement learning policy;
an output arranged to output the predicted multi-entity event and an associated human-understandable explanation comprising the domain specific language functions assigned to edges in a volume of interest of the video depicting the multi-entity event.

### Claim 2 (depends on 1)

The image processing system of claim 1 wherein the predicted multi-entity event is sent to a control system and triggers control of an apparatus selected from one or more of: an automated vehicle, a content distribution apparatus, a security apparatus, a sports training apparatus, a disaster recovery apparatus, a super-resolution apparatus, a video codec apparatus, an information retrieval apparatus, a digital personal assistant, a video collation apparatus.

### Claim 3 (depends on 2)

The image processing system of claim 2 wherein the output comprises a user interface adapted to receive input from a human in response to presentation of the human-understandable explanation so as to enable a human-in-the-loop deployment of the image processing system.

### Claim 4 (depends on 1)

The image processing system of claim 1 wherein the memory stores an automaton and wherein the processor is configured to make the selection using both the automaton and the reinforcement learning policy.

### Claim 5 (depends on 1)

The image processing system of claim 1 wherein the processor is configured, for individual ones of the edges of the graph, to select a binary label which indicates whether the edge is to be assigned a domain specific language function.

### Claim 6 (depends on 1)

The image processing system of claim 1 wherein the processor is configured to assign domain specific language functions to edges of the graph on a per-frame basis for individual frames of the video represented in the graph.

### Claim 7 (depends on 1)

The image processing system of claim 1 wherein the trained machine learning system is configured to compute a volume of interest of the video, the volume of interest depicting the multi-entity event.

### Claim 8 (depends on 1)

The image processing system of claim 1 wherein the trained machine learning system comprises a graph long short-term memory neural network.

### Claim 9 (depends on 1)

The image processing system of claim 1 wherein values of parameters of the reinforcement learning policy have been obtained by training using training data comprising volumes of interest of videos depicting known multi-agent activities.

### Claim 10 (depends on 1)

The image processing system of claim 1 wherein the trained machine learning system is a supervised machine learning system having been trained using training data comprising volumes of interest of videos depicting known multi-agent activities.

### Claim 11 (depends on 1)

The image processing system of claim 1 wherein the trained machine learning system is trained separately from the reinforcement learning policy.

### Claim 12 (depends on 1)

The image processing system of claim 1 wherein the trained machine learning system is trained jointly with the reinforcement learning policy.

### Claim 13 (depends on 1)

The image processing system of claim 1 wherein the graph formation module forms, for individual frames of the video, a grid of nodes, each node representing a patch of pixels of the video frame and connected to its nearest neighbor nodes in the grid representing the individual frame and in the grids representing the preceding and subsequent frame.

### Claim 14

A computer-implemented method for recognizing multi-entity activities, the method comprising:
storing, at a memory, a video depicting a multi-entity event, a trained reinforcement learning policy and a plurality of domain specific language functions;
computing a representation of the video as a graph of nodes connected by edges;
operating a trained machine learning system to recognize entities depicted in the video and recognize attributes of the entities, and to compute a predicted multi-entity event depicted in the video;
adding labels to the nodes of the graph according to the recognized entities and attributes;
for individual ones of the edges of the graph, selecting a domain specific language function from the plurality of domain specific language functions and assigning it to the edge, the selection being made according to the reinforcement learning policy;
outputting the predicted multi-entity event and an associated human-understandable explanation comprising the domain specific language functions assigned to edges in a volume of interest of the video depicting the multi-entity event.

### Claim 15 (depends on 14)

The computer-implemented method of claim 14 comprising sending the predicted multi-entity event and the explanation to a control system and triggering control of an apparatus selected from one or more of: an automated vehicle, a content distribution apparatus, a security apparatus, a sports training apparatus, a disaster recovery apparatus, a super-resolution apparatus, a video codec apparatus, an information retrieval apparatus, a digital personal assistant, a video collation apparatus.

### Claim 16 (depends on 15)

The computer-implemented method of claim 15 comprising displaying the explanation to a user and receiving in response user input influencing operation of the control system.

### Claim 17 (depends on 14)

The computer-implemented method of claim 14 comprising storing an automaton and making the selection using both the automaton and the reinforcement learning policy.

### Claim 18 (depends on 14)

The computer-implemented method of claim 14 comprising, for individual ones of the edges of the graph, selecting a binary label which indicates whether the edge is to be assigned a domain specific language function.

### Claim 19 (depends on 14)

The computer-implemented method of claim 14 comprising obtaining values of parameters of the reinforcement learning policy by training using training data comprising volumes of interest of videos depicting known multi-agent activities.

### Claim 20

A computer-implemented method of training an image processing system for recognizing multi-entity activities and generating human-understandable explanations of the recognized multi-entity activities, the method comprising:
storing, at a memory, a reinforcement learning policy and a plurality of domain specific language functions;
accessing training data comprising videos depicting multi-agent activities, each video labelled as depicting a particular multi-agent event of a plurality of possible multi-agent activities, and wherein entities and attributes of the entities depicted in the videos are known;
for individual ones of the videos in the training data:
computing a representation of the video as a graph of nodes connected by edges and forming a machine learning model using the graph of nodes;
using supervised machine learning to train the machine learning model to compute a predicted multi-entity event depicted in the video;
adding labels to the nodes of the graph according to the recognized entities and attributes;
using reinforcement learning to update the reinforcement learning policy.
