Huawei is transitioning from the Mate10 to enterprise intelligence, stepping into a new era of artificial intelligence. Recently, Google launched the AVA (Atomic Visual Actions) database, a collection of YouTube videos that can be annotated with 80 atomic actions like walking, shaking hands, or kicking. Each action is tagged with spatial and temporal information, totaling 57,600 video clips, 96,000 human actions, and 210,000 action labels. Google claims this dataset will enhance its video analysis capabilities for years to come, while also helping advertisers better target consumers. Their ultimate goal is to teach computers social intelligence—understanding what humans are doing, what they might do next, and their intentions.
Currently, many domestic research institutes and companies are investing heavily in behavior recognition to address the challenges of understanding video surveillance data. However, achieving the goal of machines truly "understanding" human actions requires significant effort. "Most current behavior recognition systems are still at the motion recognition stage," said Professor Li Jinping from Jinan University's Pattern Recognition and Intelligent Systems Institute. "Recognizing human actions remains a major challenge for practical applications." He added that while it’s possible for machines to detect movement, fully understanding it is still a "may have" scenario.
In terms of application, behavior recognition has shown great potential in security and other fields. For instance, in elderly care homes, a system could alert staff if an individual falls, even contacting the hospital directly. The technology is often compared to the high-tech tools seen in movies like *Mission Impossible*. Gait recognition, which identifies people based on their walking style, is another key area. It plays a crucial role in security systems as the final line of defense.
Professor Wu Fei from Zhejiang University’s Institute of Artificial Intelligence noted that this technology, initially used by Apple and Microsoft in gaming, is now being applied in public security, energy, banking, and healthcare. Gait recognition, though a form of identity recognition, overlaps significantly with behavior identification. It involves four steps: detecting human shape, segmenting, identifying, and tracking. The most challenging part is the research behind behavior recognition.
In recent years, gait recognition has played a vital role in missing persons searches and identifying suspects. In a "Wisdom" program, the technology successfully identified a suspect on the spot using gait. Even more impressively, it later recognized a dog’s silhouette through its movement. Huang Yongbiao, founder of Galaxy Drop Technology, emphasized that gait recognition excels in analyzing abnormal behaviors, which is a key advantage over other biometric features.
However, behavior recognition faces several challenges. One major difficulty is locating and tracking human motion in continuous video streams. Wu Fei explained that identifying the start and end points of an action in a video stream is extremely complex. Tracking the action robustly is equally difficult, especially when dealing with dynamic scenes, background changes, or camera shake. Scene recognition also poses a challenge, as the same action can vary greatly across different settings and times.
Li Jinping added that unlike static objects, actions evolve over time, introducing more uncertainty. A person's "running" image might actually be a jump. Only by observing multiple frames can we accurately determine what happened. This complexity makes it hard to apply one action's identification method to another. "Humans can recognize an action two or three times, but for machines, it requires extensive training data," he said.
Google’s AVA dataset aims to provide deeper insights into models by offering fine-grained temporal and spatial annotations. This will not only improve the accuracy of human activity recognition but also enable broader applications in the future.
Looking ahead, both equipment platforms and theoretical models are gaining momentum. Li Jinping highlighted China's strong potential in behavior recognition over the next five to ten years. He suggested focusing on developing domestic equipment and platforms, as many teams currently rely on foreign solutions. Policy support would be essential.
Wu Fei also stressed the need for stronger theoretical models and data-driven research. He pointed out that behavior recognition involves multiple states over time, with varying postures at each moment. Not all frames carry equal importance, and different joints may contribute differently to motion discrimination. These complexities require advancements in fundamental theories and models.
This series is an effect maker on the stage, including laser lights with multi-segment shape changes, electrodeless lights with sudden changes in color angles, matrix lights that can display fixed patterns or letters, etc., both in color and light shape. outstanding effect. The laser light has been edited to show a 3D stereoscopic effect. With the delay effect in the venue, the laser light with strong penetrating power can show a more beautiful effect. If you want the stage to be more cool, the effect light series is also an essential option.
Stage Effect And Laser Lights,Laser Light,Disco Laser Light,Sky Laser Light
Guangzhou Cheng Wen Photoelectric Technology Co., Ltd. , https://www.cwdisplay.com