AI-enabled Real-time Aerial Target Detection and Data Transport

CrowdAI specializes in AI/ML for automated target recognition (ATR) on full motion video (FMV). With years of DOD past performance in this mission space (DON, JAIC, MAVEN, USAF), CrowdAI will develop an end-to-end computer vision (CV) system for DON aerial platforms. It will ingest and process video, identify targets, and serve video snippets to onboard sensor operators with automated alerts to their multi-function displays. CrowdAI was the first to develop segmentation models on FMV, and today’s advanced ATR models can detect and identify surface contacts in video. Snippets will be compressed automatically, reducing onboard storage and facilitating immediate backhaul via TCDL. The final system will run in the background and support multiple DON missions. CrowdAI intends to work with Pixspan, the unparalleled expert in lossless compression, to develop the full solution. We will design a representative system that includes an ATR model, a snippet export tool, and the functionality to receive, compress, forward media using the fastest possible upload. For fine-tuning, the system will also include a feedback mechanism to augment training data in CrowdAI’s no-code CV software.  We will add demonstrable improvements to real-time CV analysis and the onboard curation, efficient download, and rapid backhaul of high-resolution data. Our software supports intuitive review and retraining of CV at the tactical edge, providing an end-to-end system that not only continuously improves, but is serviceable by DON personnel without any training in data science or AI/ML. Developed from TRL-9 commercial technologies, DON will be able operate this system with its existing workforce without additional personnel. The system will act as a force multiplier, relieving analytic burden on operators to pursue more advanced tasks.   In the face of a constrained compute environment, the project will investigate the feasibility of dynamic reallocation of GPU/CPU resources to prioritize specific processes. That is, ATR for threat detection may require real-time (sub-second) processing, while search and rescue can accept processing every second. The ability to adjust compute allocation will ensure that more creew needs are met. We will interview stakeholders to identify other features beneficial to their missions for development. As experts in our respective fields, we have the technologies and workforce to rapidly develop and implement this solution while exceeding expectations. CrowdAI is actively providing trusted AI/ML for FMV to both the DON and U.S. Air Force. While this mission set and the unique compute environment are novel, both aspects of this project are squarely within our respective areas of expertise.  Lastly, for an interim project to bridge Phase I and Phase 2, CrowdAI suggests the curation of video data from the DON operations platform and to begin labeling data. This process can add significant lead time to a project. Generally, it is low cost, but high impact.