NV300 is an ultra-small-form-factor (USFF) MIL-SPEC rugged edge compute solution. Designed using a Modular Open Systems Approach (MOSA) with a modular chassis and architecture design, integrating either the NVIDIA Jetson AGX Orin edge AI system on module (SOM), NV300 provides the performance and flexibility demanded by emerging autonomous and unscrewed missions, in an ultra-compact and lightweight form factor. NV300 is engineered to withstand austere environments and is SWaP-optimized for integration and deployment in highly space-constrained platforms.

Edge AI Platform	SFF & SWaP	Open Modular Architecture	MIL-STD 810 Rugged
Integrated with the NVIDIA Jetson AGX Orin (32GB or 64GB), delivering up to 275 TOPS of AI performance for next-gen autonomous and uncrewed mission-critical applications.	Ultra-compact and lightweight, SWaP-optimized for integration and deployment in highly space-constrained vehicles.	Jetson AGX Orin and SMARC x86-based architecture/processor options. Robust IO and configuration options, including multiple USB 3.0, GbE, serial, display, video capture, GP, GPIO.	Engineered for austere environments. -40°C to +55 °C operating temp, passive cooled. Fully sealed. MIL-STD-810H, MIL-STD-461G, MIL-STD-1275E/704F.

SDI (Serial Digital Interface)

AI targeting systems leverage data from computer sensors and various sources to statistically assess potential targets, aggregating information from drone footage, surveillance data, open-source intelligence, and monitored movements and behaviors. This comprehensive approach enables efficient target detection and tracking. Supporting these capabilities, a state-of-the-art embedded hardware system empowers small flying robots to perform real-time onboard computation necessary for object tracking.

In computer vision applications, video processing and display are essential tasks, requiring seamless transmission of video or images from cameras to the host vision system. To achieve this, system designers must select the appropriate video interface standard based on factors like data rates, cable lengths, and power delivery options. HD-SDI and 3G-SDI are widely adopted standards in professional video production and broadcasting environments, supporting resolutions up to 1080p at 30/60Hz and offering data rates up to 3 Gbps. These signals are transmitted via a single 75Ω coaxial cable, making them suitable for applications requiring high fidelity and real-time video transmission, with cable lengths of up to 200m (using RG59/RG6 cables).

Furthermore, the system allows connection to 4 video input channels, including 4 Channel HD-SDI video channels or 4 Channel 3G-SDI Video Input plus 2 channel SDI sources video output. It generates from 2 up to 4 video output channels while maintaining low latency between input and output channels. The generated output channel provides a Bird's-Eye-View created from 4 SDI input channels, with each output channel selectable into one main channel for enhanced flexibility and functionality.

Detecting Objects in Point Clouds with NVIDIA CUDA-PointPillars

(Sourced from NVIDIA) 

NV300 can generate a point cloud, a data set consisting of points within a coordinate system. Each point in a point cloud holds valuable information, such as three-dimensional coordinates (X, Y, Z), color, classification value, intensity value, and time. Point clouds are primarily produced by 3D Lidar systems, which are extensively used in various NVIDIA Jetson applications, including autonomous driving, perception modules, and 3D modeling. A key application of point clouds is utilizing long-range and high-precision data sets to achieve 3D object detection, which is essential for perception, mapping, and localization algorithms.

 

GPU acceleration of object detection on video stream using CUDA

NV300 excels in object detection, a crucial technology for applications such as autonomous driving and intelligent video analytics. Achieving high accuracy in object detection necessitates extensive training with large datasets, a task ideally suited for the parallel computing capabilities of NVIDIA GPUs. These GPUs are specifically designed to efficiently train large neural networks, enabling the rapid and effective generation of datasets for object detection inference.

Object detection plays a pivotal role in computer vision, as well as in image and video processing. Achieving top accuracy and fast detection under varying conditions—such as changes in position, scale, illumination, and noise—poses a significant challenge that cannot be effectively addressed by sequential processing on a single-core General Purpose Central Processing Unit (GPCPU). This post outlines the essentials for leveraging NVIDIA GPUs to implement high-performance object detection pipelines quickly and efficiently, addressing the inherent complexities of real-time object detection in dynamic environments.

(Sourced from NVIDIA)

System Diagram

 

 

 

 

Block Diagram

Appearance

Optional:

Dimension