Abstract
A moving observer must determine her direction of motion and detect moving objects. Once heading is known, a moving object can be detected by examining local image motion differences and identifying those that have large angular differences from the radial optic flow pattern or that have a larger than average magnitude. A physiologically based model for computing heading (Royden, 1997) performs a motion subtraction using operators based on physiological properties of cells in the Middle Temporal visual area (MT). This model computes heading well. Here, we tested the model's ability to detect moving objects in the scene. The model was modified to detect potential moving object borders based on an increased response magnitude compared to the average response across the scene or a preferred direction of motion that differs significantly from the radial pattern expected for the computed heading. We simulated observer translational motion of 200 cm/sec toward the center of two frontoparallel planes, 400 and 1000 cm from the observer. A 6x6 deg moving object moved leftward at a speed of 7.5 deg/sec. The horizontal and vertical positions of the object were varied between -7, 0 and 7 deg. Speed and angle thresholds for signaling a possible moving object were determined empirically, to give the maximum object detection with the minimum false positives. Once these thresholds were established, they were held constant for all conditions tested. We found that in most cases the model detected the borders of the moving object very well, with an average of 10.4 out of 12 of the operators located along the object edges signaling a moving object border and 2.4 of 157 operators outside the object edges giving false positives. The results were similar when tested with the addition of observer rotation of 5 deg/sec about the X, Y or Z axes.
Supported by NSF Grant #IBN-0343825