Object Tracking

Deep Vision's data abstraction technology delivers real-time object tracking.

Deep Vision's novel concept of data abstraction quickly transforms abundant sensor data into a form that is easily classified and efficiently analysed.

The abstractions created from the raw data, coupled with their position on the sensor and relative positions over time, provide the foundation for real-time object tracking.

The abstractions are not dependent on the sensing platform, type of sensor or other objects within the sensor's data. Therefore, the abstractions, and by extension the objects they represent, can be tracked independent of sensor modality, sensor motion, and object motion.

With and without prior knowledge of the target objects, or their characteristics, Deep Vision's object tracking technology delivers a highly robust, real-time target locking solution.

Exploitation Value
  • Operator Assistance
  • Target locking
  • Motion Analysis and Prediction

Features

  • Real-time object tracking
  • Simultaneous tracking of multiple objects
  • Operates in complex and cluttered environments
  • Context-based motion analysis
  • Sensor and object independent
  • Moving platform and moving target
  • Designation of a specific target whilst maintaining awareness of all others.
  • Target designation in one sensor type enables detection in disparate sensor types.
Operating Facts
  • Operating System: Any (GNU/Linux recommended)
  • Hardware Requirements: None
  • Sensor Modalities: Visual, Thermal, Sonar
  • Timings†: 100+ FPS
  • Runtime Memory Requirements†: 300 KB
  • Storage Requirements‡: 1.1 KB

† Typical. Based on a 640 x 480 data set
‡ Typical. Based on 45 abstractions (avg. 25 symbols each)

Input Requirements

  • Real-time acquisition from visual, thermal, or sonar sensors.
  • Archived video and imagery.

Examples

Evading Target

An evading human target being chased down in the woods. Real-time tracking in a cluttered natural environment with chaotic sensor motion. The target moves unpredictably through a wooded area, leaving and entering the field of view. The target is continuously reacquired as it re-enters, even when it is partially obscured and only visible for a fraction of a second. This video was recorded using legacy interlacing - very noisy.

Another evading human target being chased down in the woods but this time recorded using progressive scan. Note how extreme sensor motion, changes in perspective and scale and wildly unpredictable behaviour of the target, including obsfucation to the point of loss, has little affect, if any, on tracking capability.



Moving Target Moving Platform

In the maritime setting, a person in distress is robustly tracked and monitored, independent of the scale of the person, sea state, ocean and platform motion. In this instance it's a helicopter but it could just as easily be a UAV on a scouting mission.


The user selects the target to track, and tracking of that target will continue as long as the object remains within the field of view. Tracking is unaffected by target/platform motion, environmental effects (e.g. rain, sea spray), target scale, perspective, changing view or occlusion.


Designation of a fleeing car from the dash camera on a police vehicle. The car is tracked during the pursuit, demonstrating robust tracking in the scenario of moving target/moving platform.


Designation and tracking of a human target. The target is tracked whilst the camera experiences high degrees of motion. The tracking is independent of target/platform motion, scale/pose of the target, in addition to changing target/background composition and dynamic reacquisition.



Camouflaged and Slow Moving Targets

The operator designates a target at runtime and that target is tracked within the field of view, independent of target/platform motion, target scale/pose, and target/background composition. Note the extent of camouflage and track persistence through loss of focus.


Tracking Across Sensor Modalities

These videos demonstrate the use of Deep Vision's sensor exploitation technology for use in tracking of people regardless of sensor modality (e.g. Visual and/or IR) and the continual target monitoring by all assets in the same area.

Designation of a person in one sensor and handover that target for acquisition by another. Target information is immediately transmitted to all members of the distributed system.


Designation of a person in one area and reacquisition of the target in another area by another sensor. Target information is immediately transmitted to all members of the distributed system.


Aerial Tracking of Ground Vehicles

These videos demonstrate the use of Deep Vision's sensor exploitation technology for use in aerial tracking of ground vehicles regardless of sensor modality (e.g. Visual and/or IR).

The user designates a ground vehicle. That vehicle is tracked whilst maintaining awareness of all other similar vehicles in the scene.


The user designates a ground vehicle. That vehicle is tracked whilst maintaining awareness of all other similar vehicles in the scene.


Motion Analysis

These videos illustrate the use of Deep Vision's sensor exploitation technology with low-level motion analysis.

Simultaneous tracking of two people in motion (delineated by bounding box).


Tracking of an intermittently moving person. The person remains locked even when motionless (indicated by the constant colour of the bounding box). Note that the pick-up truck near the end of the sequence (top right corner) is ignored, demonstrating the ability to differentiate between items of interest in any scene.


Preferential tracking of a moving person. The moving vehicle is ignored in favour of the person even after the person has left the scene.


Dynamic Knowledge Base

These videos illustrate the use of Deep Vision's sensor exploitation technology with a dynamic knowledge base. The user selects and labels an object from within the scene. The description and label pair is added to the knowledge base.

The user selects two objects: a cross and the letter H. Each object is given a meaningful label: First Aid and Hospital respectively. These objects are then placed in the environment and the camera is moved within it. Each object presented on the sensor and found in the knowledge base is highlighted and labelled accordingly.


The users selects three objects: the letter H, an airplane and a cross. Each object is given a meaningful label: Hospital, Airport and First Aid respectively. These objects are then moved throughout the environment.