Summary:
UCLA Researchers in the Department of Electrical and Computer Engineering have developed a new computationally efficient algorithm that detects and tracks motion in videos.
Background:
Computer vision and video processing have many applications in visual security and surveillance such as people counting, action recognition, anomaly detection, smart environments, and activity localization and tracking. These video applications all require pre-processing steps such as video dynamic (motion) processing or more long-term processing such as the appearance/disappearance of objects; such motion change detection events are pre-processing steps for subsequent identification, classification, or prediction tasks. Currently, the algorithms used for motion and change detection are computationally heavy, require training with large known datasets, and are sensitive to sudden illumination changes and environmental conditions, background motion, shadows, and camouflage effects. There remains an unmet need for a single algorithm that simultaneously addresses all key challenges that can accompany real-world video processing.
Innovation:
Professor Jalali’s Laboratory has developed PhyMotion, a novel method for video motion and event detection that integrates image sensing with a computer vision sensing algorithm. The algorithm runs on a digital processor that produces an output for visualization, which is then used for storage and subsequent detection and classification. It converts a conventional image sensor to an event-based sensor. This method is based on Jalali Lab’s physics-inspired approach to design of computational imaging algorithms known as PhyCV (physics inspired computer vision). The approach was inspired by physical processes behind the Photonic Time Stretch, a data acquisition method invented by the laboratory that has been the most successful method for capturing, digitizing, and analyzing of ultrafast physical signals in real-time. The new algorithm, called PhyMotion operates on both grayscale or color images and can identify and isolate motion of each color channel. Video processing for both fully-loaded videos and real-time video monitoring are computationally accelerated through various means, allowing for quicker processing with less computational power. The proposed innovation thus overcomes previous shortcomings of computer vision and video processing technologies, with versatile applicability in fields such as security, medical diagnostics and autonomous vehicle operations.
Potential Applications:
• Structural health monitoring
• Autonomous driving
• Security and defense
• Human-computer interface (eye movement detection)
• Vibration monitoring
• Medical diagnostics
• Liquid tracking
• Human tracking
• Vibration monitoring
• Data Compression
• Metal process monitoring
Advantages:
• Operational on grayscale and color videos
• Real-time analysis
• Low-light enhancement
• Works with conventional images sensors
Development to Date:
Algorithm developed and successfully demonstrated.
Reference:
UCLA Case No. 2023-131
Lead Inventor:
Bahram Jalali (Laboratory: www.photonics.ucla.edu)