Showing 1–3 of 3 results for author: Morris, D D
-
A Pyramid CNN for Dense-Leaves Segmentation
Authors:
Daniel D. Morris
Abstract:
Automatic detection and segmentation of overlapping leaves in dense foliage can be a difficult task, particularly for leaves with strong textures and high occlusions. We present Dense-Leaves, an image dataset with ground truth segmentation labels that can be used to train and quantify algorithms for leaf segmentation in the wild. We also propose a pyramid convolutional neural network with multi-sc…
▽ More
Automatic detection and segmentation of overlapping leaves in dense foliage can be a difficult task, particularly for leaves with strong textures and high occlusions. We present Dense-Leaves, an image dataset with ground truth segmentation labels that can be used to train and quantify algorithms for leaf segmentation in the wild. We also propose a pyramid convolutional neural network with multi-scale predictions that detects and discriminates leaf boundaries from interior textures. Using these detected boundaries, closed-contour boundaries around individual leaves are estimated with a watershed-based algorithm. The result is an instance segmenter for dense leaves. Promising segmentation results for leaves in dense foliage are obtained.
△ Less
Submitted 4 April, 2018;
originally announced April 2018.
-
A View-Dependent Adaptive Matched Filter for LADAR-Based Vehicle Tracking
Authors:
Daniel D. Morris,
Regis Hoffman,
Paul Haley
Abstract:
LADARs mounted on mobile platforms produce a wealth of precise range data on the surrounding objects and vehicles. The challenge we address is to infer from these raw LADAR data the location and orientation of nearby vehicles. We propose a novel view-dependent adaptive matched filter for obtaining fast and precise measurements of target vehicle pose. We derive an analytic expression for the matchi…
▽ More
LADARs mounted on mobile platforms produce a wealth of precise range data on the surrounding objects and vehicles. The challenge we address is to infer from these raw LADAR data the location and orientation of nearby vehicles. We propose a novel view-dependent adaptive matched filter for obtaining fast and precise measurements of target vehicle pose. We derive an analytic expression for the matching function which we optimize to obtain target pose and size. Our algorithm is fast, robust and simple to implement compared to other methods. When used as the measurement component of a tracker on an autonomous ground vehicle, we are able to track in excess of 50 targets at 10 Hz. Once targets are aligned using our matched filter, we use a support vector-based discriminator to distinguish vehicles from other objects. This tracker provides a key sensing component for our autonomous ground vehicles which have accumulated hundreds of miles of on-road and off-road autonomous driving.
△ Less
Submitted 25 September, 2017;
originally announced September 2017.
-
LADAR-Based Mover Detection from Moving Vehicles
Authors:
Daniel D. Morris,
Brian Colonna,
Paul Haley
Abstract:
Detecting moving vehicles and people is crucial for safe operation of UGVs but is challenging in cluttered, real world environments. We propose a registration technique that enables objects to be robustly matched and tracked, and hence movers to be detected even in high clutter. Range data are acquired using a 2D scanning Ladar from a moving platform. These are automatically clustered into objects…
▽ More
Detecting moving vehicles and people is crucial for safe operation of UGVs but is challenging in cluttered, real world environments. We propose a registration technique that enables objects to be robustly matched and tracked, and hence movers to be detected even in high clutter. Range data are acquired using a 2D scanning Ladar from a moving platform. These are automatically clustered into objects and modeled using a surface density function. A Bhattacharya similarity is optimized to register subsequent views of each object enabling good discrimination and tracking, and hence mover detection.
△ Less
Submitted 25 September, 2017;
originally announced September 2017.
