- They considered there are two major uses of gait recognition:
- Gait recognition is non intrusive and could be used in a distance which might contribute to identifying criminals
- Gait recognition also has a medical application which could aid doctors and trainers identify changes in walking pattern
- No one has yet came up with a model that is both accurate and marketable
- There are two main approaches of to gait recognition research
- Automatic analysis of video imagery from multiple cameras
- Radar technology to record gait cycle
- Given that humans can identify others by their manner of walking, there had been some computer vision algorithms developed for people identification and activity classification
- Derive the silhouette of the person walking:
- Use background subtraction algorithm to single out the figure
- Ignore foreground color to remove the effects of clothing
- Use only scale-normalized binary silhouette
- Use the silhouette to construct a set of moment related figures and those are used to identify a person
- Considered distance to be the key benefit of gait recognition.
- One major application they see of gait recognition is airport terminals where distances made facial recognition difficult while metrics like torsal proportions and relations are easier to detect.
- Use two main approaches:
- Computer analysis to parse digital video into "four-dimensional walk vectors," "cross-conditioning mapping methods", and "time-normalized, joint-angle trajectories in the walking plane"
- Radar to detect speed of limb movements and radar has the advantage of not obstructed by light and weather condition.
Machine Learning Scientist
Depending on what works best, we will use either LEDs or colored patches worn on the body to simplify the work done in analyzing film of the user.
To gather additional data on the movement of the person being authenticated, we will also use inertial sensors worn on the body. We are considering a combined accelerometer-gyroscope such as the MPU-3050, an IMU such as the MinIMU-9 v5, or just combining an individual accelerometer and gyroscope.
In order to simplify the authentication process, we will place the sensors, markers and data transmitter on some sort of harness to make it easy to take on and off.
To gather data for authentication, we will use a video camera and markers worn on the body to analyze movement. One or at most two cameras will be necessary, since we can limit the motion of the user, for example by instructing them to walk toward the camera. We will prioritize frame rate over resolution since the markers should be easily detectable.
We are considering using a Machine Learning library such as weka to perform gait analysis and pattern matching for authentication. We may also write our own software in order to better meet the needs of our project.
To transmit data from the sensor apparatus, we will use a Raspberry Pi to gather data from the sensors and transmit it to be analyzed.
- Camera must be able to pick out the distinct points on the subject
- Accelerometer/gyroscope must be able to pick up data and transmit them to the processing unit
Requirements: Non Functional
- The camera data must be clear and distinct enough that the computer can establish the distinct points without confusion or mistakes
- The accelerometer/gyroscope data must be reasonable enough so that it does not jump around with random values
- The collective data from the camera and the accelerometer/gyroscope must establish a pattern that is distinct to the subject