Event-based image feature extraction

1. A method for event-based image feature extraction comprising:

reducing an accumulated magnitude of a leaky integrate and fire (LIF) neuron based on a difference between a current time and a previous time;

receiving at least one event input from at least one dynamic vision sensor (DVS) pixel at the current time;

weighting the at least one received input;

adding the at least one weighted input to the reduced magnitude to form an accumulated magnitude of the LIF neuron at the current time; and

if the accumulated magnitude reaches a threshold, firing the neuron and decreasing the accumulated magnitude,

wherein the decrease upon firing is equal to the accumulated magnitude if the accumulated magnitude is less than a base, or equal to the base if the accumulated magnitude is greater than or equal to the base.

2. The method of claim 1 wherein each at least one received input is either +1 or −1 prior to weighting.

3. The method of claim 1 wherein the LIF neuron receives event inputs from a plurality of DVS pixels in a receptive field.

4. The method of claim 3 wherein the event inputs from the receptive field are individually weighted to implement a Hough transform.

5. The method of claim 3 wherein the receptive field includes an array of DVS pixels.

6. The method of claim 3 wherein the event inputs from the receptive field are individually weighted to implement a histogram of gradients (HOG).

7. The method of claim 1 wherein a first plurality of LIF neurons is arranged in a first layer.

8. The method of claim 7 wherein another plurality of LIF neurons is arranged in another layer.

9. The method of claim 8 wherein each of the other plurality of LIF neurons receives event inputs from at least one of the first plurality of LIF neurons.

10. The method of claim 8 wherein each of the other plurality of LIF neurons receives event inputs from all of the first plurality of LIF neurons.

11. The method of claim 8 wherein each of the other plurality of LIF neurons triggers an event output upon reaching an accumulated threshold based on an event input from any of the first plurality of LIF neurons in its receptive field.

12. The method of claim 8 wherein each of the other plurality of LIF neurons receives event inputs from all of the first plurality of LIF neurons in its receptive field, but triggers an event output only upon an event input from a corresponding one of the first plurality of LIF neurons.

13. The method of claim 1 wherein the reduction is based on an exponential decay function.

14. The method of claim 1 wherein the LIF neuron is one of a plurality of LIF neurons arranged in a plurality of levels, the method further comprising:

implementing a directionally sensitive filter using a HOG in one of the plurality of levels; and

implementing a Hough transform responsive to the directionally sensitive filter in another of the plurality of levels.

15. A non-transitory computer readable medium tangibly embodying a program of instruction steps executable by a processor for extracting event-based image features through a dynamic vision sensor, the instruction steps comprising:

reducing an accumulated magnitude of a leaky integrate and fire (LIF) neuron based on a difference between a current time and a previous time;

receiving at least one event input from at least one dynamic vision sensor (DVS) pixel at the current time;

weighting the at least one received input;

adding the at least one weighted input to the reduced magnitude to form an accumulated magnitude of the LIF neuron at the current time; and

if the accumulated magnitude reaches a threshold, firing the neuron and decreasing the accumulated magnitude,

wherein the decrease upon firing is equal to the accumulated magnitude if the accumulated magnitude is less than a base, or equal to the base if the accumulated magnitude is greater than or equal to the base.

16. A method for event-based image processing comprising:

providing a plurality of leaky integrate and fire (LIF) neurons wherein each of the plurality of LIF neurons communicates with at least one other of the plurality of LIF neurons;

reducing an accumulated magnitude of each of the plurality of LIF neurons based on a difference between a current time and a previous time;

receiving by at least one of the plurality of LIF neurons at least one event input from at least one dynamic vision sensor (DVS) pixel at the current time;

weighting the at least one received input;

adding the at least one weighted input to the reduced magnitude to form an accumulated magnitude of the at least one of the plurality of LIF neurons at the current time; and

if the accumulated magnitude of the at least one of the plurality of LIF neurons reaches a threshold, firing the at least one of the plurality of LIF neurons and decreasing its accumulated magnitude,

wherein the decrease upon firing is equal to the accumulated magnitude if the accumulated magnitude is less than a base, or equal to the base if the accumulated magnitude is greater than or equal to the base.

17. The method of claim 16 , further comprising:

if the accumulated magnitude of the at least one of the plurality of LIF neurons reaches a threshold based on receiving signals from a plurality of event inputs, but is conditioned to fire only upon receiving a signal from a particular one of the plurality of event inputs, firing the at least one of the plurality of LIF neurons and decreasing its accumulated magnitude upon receiving the signal from the particularone of the plurality of event inputs.

18. The method of claim 16 wherein the plurality of leaky integrate and fire (LIF) neurons is provided in a multi-dimensional array having at least two conceptual dimensions (2D).

19. The method of claim 16 wherein:

the plurality of LIF neurons is provided in a multi-dimensional array having three conceptual dimensions (3D), the 3D array comprising a plurality of layers of LIF neurons where each LIF neuron of a layer is interconnected to all of the LIF neurons of each directly adjacent layer, and

a first of the plurality of layers implements an un-normalized histogram of gradients (HOG) and a second of the plurality of layers implements a normalized HOG.

20. The method of claim 16 wherein:

the plurality of LIF neurons is provided in a multi-dimensional array having three conceptual dimensions (3D), the 3D array comprising a plurality of layers of LIF neurons where each LIF neuron of a layer is interconnected to all of the LIF neurons of each directly adjacent layer,

a first of the plurality of layers implements a histogram of gradients (HOG) and a second of the plurality of layers implements a Hough transform,

the first of the plurality of layers is interconnected with the DVS pixels, each LIF of the first of the plurality of layers having a receptive field responsive to a plurality of the DVS pixels, and

the receptive fields are non-overlapping.