AN IMPROVED 2D OPTICAL FLOW SENSOR FOR MOTION SEGMENTATION

Vdd

Vdd

Vdd

Vdd

β∣MI
+
y∣∣δf∣∣

⅛⁴Γj≤f^ι ⅛

P

^i,j pass transistors

threshold
current

segmentation
current

vi,j+1

Fig. 3. Schematics of a discontinuity unit.

output behavior (due to the boundedness of g) if the output
conductance R is large: P_ii = 1, if the weighted measure
of the flow gradient and the brightness constraint deviation
is larger than a threshold a, and P_ij = 0 otherwise.

Closing the feedback loop, the two relatively simple net-
work stages of the system solve a typical combinatorial
problem, which is computationally hard. Unlike in other
network solutions of such problems [6], the network ar-
chitecture is non-homogeneous. The discontinuity network
thereby performs a typical line process [8], although it re-
mains fully deterministic. Hence, the found solutions might
be only sub-optimal which can be reflected by hysteretic be-
havior in the activity of the discontinuity network.

3. HARDWARE AVLSI IMPLEMENTATION

An 11x11 array of the described optical flow system has
been implemented in a double-poly double-metal 0.8 m
BiCMOS process. Each pixel consists of an optical flow
unit plus two discontinuity units. The schematics of the op-
tical flow unit is basically as reported in [1] although im-
proved [7]. The estimated optical flow field is encoded as
the continuous voltage distributions U_ij and V_ij in two re-
sistive layers with respect to some reference potential, where
the output signal range is roughly ±0.5 V. The optical flow
units can reliably report the speed of visual motion over al-
most 3 orders of magnitude.

The schematics ofa single discontinuity unit are shown
in Figure 3. The circuit approximates the dynamics (4) with
the output P_ij being inverted. The error measures (∣∣∆n∣∣
and ∣∣ΔF∣∣) are implemented by bump circuits [9] that pro-
vide the local segmentation current accordingly. The out-
put of the discontinuity units controls a pair of pass transis-
tors sitting in between two neighboring units of the optical
flow network in order to break the lateral conductances or
to leave them at some preset value .

In total, a single pixel consists of roughly 200 active el-
ements, occupying a chip area of (170 μm)². A substantial
fraction of this area, however, is used for all the nearest-
neighbor connections of the different signals. The fill-factor
is at low 4% and power consumption is 80μWφixel in steady-
state.

Fig. 4. Detecting motion discontinuities. The scanned se-
quence of the chip’s output while seeing a dark dot on a
light background, moving from the left upper to the right
lower corner of its visual field.

4. RESULTS

We waive a detailed characterization of the optical flow units
(which can be found in [7]) and report instead the response
of the complete system in two visual experiments, performed
under real-world conditions.

In the first example, the chip was presented with a stim-
ulus consisting of a dark moving dot on a light background.
Figure 4 shows the sampled responses of the chip while the
dot was moving from the upper left to the lower right cor-
ner of its visual field. The estimated optical flow field is
shown superimposed onto the images of the photorecep-
tor output, while the associated activity of the discontinu-
ity units (P and Q) is displayed as binary images below
each frame. Note that the activity pattern of the disconti-
nuity units approximately reflects the contour of the dark
dot. However, the chip has difficulties to achieve a closed
contour that completely separates figure and background.
Nevertheless, the optical flow estimate is improved insofar
as it predominantly preserves a sharp flow gradient at the
dot’s outline.

II - 334