การติดตามเส้นทางการ1.1. Previous Work
There exists a large body of literature related to our approach.
Only the most relevant ones are briefly reported
here. Many approaches [3,7] in motion detection and tracking
assume that the scene can be approximated by a plane
(i.e. a flat ground), and image transformation is modeled as
a 2D affine or projective transformation. These approaches
work well in general, even when the reference plane corresponding
to the affine stabilization does not correspond to
the ground plane. However, when this assumption is violated
due to the presence of tall structures such as buildings,
or trees, the observed motion of these stationary structures
do not follow the planar approximation of the scene, and
these structures are often incorrectly detected as independently
moving regions.
In case of strong parallax, geometric and shape constraints
are integrated for distinguishing independently
moving regions from parallax regions. In [9], the mismatch
between a spatiotemporal gradient measurement and projected
flow is used as the indicator for independent motion.
Whereas, a constant reference plane is assumed and adaptive
thresholding is needed for pixel classification. [13] enforces
the trifocal constraints in image triplets and clusters
the tracked image features (e.g. points) into different motion
groups. However, the motion segmentation is only applied
to sparsely matched features.
Many approaches are based on the “Plane+Parallax” representation
[5, 6, 8, 10–12]. [5] proposed a pair-wise rigidity
constraint in multiple frames which relates the relative
depth and parallax displacements. In [11], a shape constraint
for each point triplet is proposed based on the trifocal
constraint by eliminating the relative depth. All these methods
require that the reference plane is constant and therefore
the distances of 3D points to the plane are constant. In contrast,
our method is capable of handling different reference
planes, and does not require the manual selection of the reference
planeเครื่องที่ของวัตถุบนระนาบด้วยกล้อง
การติดตามเส้นทางการ1.1. Previous WorkThere exists a large body of literature related to our approach.Only the most relevant ones are briefly reportedhere. Many approaches [3,7] in motion detection and trackingassume that the scene can be approximated by a plane(i.e. a flat ground), and image transformation is modeled asa 2D affine or projective transformation. These approacheswork well in general, even when the reference plane correspondingto the affine stabilization does not correspond tothe ground plane. However, when this assumption is violateddue to the presence of tall structures such as buildings,or trees, the observed motion of these stationary structuresdo not follow the planar approximation of the scene, andthese structures are often incorrectly detected as independentlymoving regions.In case of strong parallax, geometric and shape constraintsare integrated for distinguishing independentlymoving regions from parallax regions. In [9], the mismatchbetween a spatiotemporal gradient measurement and projectedflow is used as the indicator for independent motion.Whereas, a constant reference plane is assumed and adaptivethresholding is needed for pixel classification. [13] enforcesthe trifocal constraints in image triplets and clustersthe tracked image features (e.g. points) into different motiongroups. However, the motion segmentation is only appliedการจับคู่เบาบางคุณลักษณะหลายวิธีขึ้นอยู่กับการแสดง "เครื่องบิน + Parallax"[5, 6, 8, 10-12] [5] เสนอความแข็งแกร่งที่ pair-wiseข้อจำกัดในกรอบที่เกี่ยวข้องสัมพันธ์ความลึกและ parallax ที่ displacements ใน [11], ข้อจำกัดของรูปร่างแต่ละจุด triplet จะเสนอตาม trifocalข้อจำกัด โดยการตัดความสัมพันธ์ลึก วิธีการเหล่านี้ต้องการให้ระนาบอ้างอิงเป็นค่าคงดังนั้นระยะทางของเครื่องบิน 3D จุดมีค่าคง ความเปรียบต่างวิธีการของเรามีความสามารถในการจัดการอ้างอิงอื่นเครื่องบิน และไม่ต้องการเลือกการอ้างอิงด้วยตนเองplaneเครื่องที่ของวัตถุบนระนาบด้วยกล้อง
การแปล กรุณารอสักครู่..
