Bottlenose™ is a family of high resolution smart cameras for industrial automation and robotics applications. It is available in monocular and stereo versions. The stereo version includes dual image sensors, and both monocular and stereo versions have hardware synchronized triple-axis gyroscopes, accelerometers, and magnetometers. Bottlenose™ can be connected over standard Ethernet using GigE Vision 2.1. It has on-camera simultaneous processing for HDR, feature point detection and matching, dense disparity, and AI.
Disparity is used to calculate differences between a left image and a right image for determining depth. On Bottlenose™ this process begins by capturing synchronized images and then undistorting and rectifying them. Each pixel in the two images is compared to find
the nearest match. Labforge uses a version of the well-known Semi-Global Matching (SGM) algorithm. Bottlenose™ computes this at up-to 4K resolution with a 200+ MP/s internal throughput. This groundbreaking performance allows faster depth, wider angles, and/or longer range coverage. Frame rate is limited by user configuration, settings, and ultimately by the 1Gigabit/s output over Ethernet.
Results are output as rectified disparity images or a fully triangulated dense point cloud.
In a traditional setup, this level of 3D disparity generation would be done on an external host PC with a very large GPU or FPGA. Bottlenose™ does not need a large PC for operation.
Feature detection is a method for fingerprinting interesting areas in an image. Bottlenose™ computes corner points in the left and right images, and also corner points in images from a previous time step. It detects points using FAST and GFTT. Built-in hardware acceleration is used to calculate A-KAZE descriptors for each point, and these are output as GigE Vision’s chunk-data or contours. Optionally, these points can also be output as a fully triangulated sparse 3D point cloud.
Customers can use this front-end processing result to build their own pick & place, automated harvesting, SLAM, navigation, and obstacle avoidance solutions.
Bottlenose™ With Hardware Acceleration On-Camera Processing
AI in imaging applications usually requires the use of deep neural networks. Bottlenose™ has hardware acceleration for AI inference for neural network models for example YOLOv3, SSD, and others. Labforge provides pre-trained models that run out of the box. Users can also load their own models into the camera, provided that the layers are supported. Dataset collection, labelling, and model training services are available as part of support.
Bottlenose™ has 4 GB of on-board LP-DDR4 camera and 320 MB of flash storage to cater to large model sizes. This allows on-camera computation of full scale neural networks without the need for quantization and pruning.
Applications include segmentation, classification, detection, pose detection, and monocular depth estimation. These results are output over GigE Vision 2.1 and can be read via chunk data or as contours.
HDR combines low, mid, and high exposure image frames to create images with larger bit-depth. Bottlenose™ has hardware acceleration for HDR and local tone mapping. This allows Bottlenose™ to excel in simultaneous low light and bright light situations. These computations are done on-camera.
HDR-processed images can be acquired from Bottlenose™ using GigE Vision 2.1. Frame rate is limited by user configuration, settings, and ultimately by the 1Gigbit/s output over Ethernet.
Contact
DEMO?
Interested in our products? Feel free to contact us. We would love to hear from you.
ADDRESS
103 Bauer Pl, Suite 6,
Waterloo, ON, Canada
N2L 6B5
CONTACT
contact@labforge.ca
Tel +1-226-929-7740