Real-Time Human Action Recognition Using Deep Learning

2. Related Work

Researchers have devoted a lot of efforts in the field of artificial intelligence that studies the recognition of human activity through videos, and in recent decades they have found many different ways (Lei 2019), (Yadav 2019) (Figure 1) to improve prediction results in terms of motion recognition accuracy and recognition speed in live video clips. Some clip-to- skeleton techniques were used (Parisi 2020) to discover differences by creating a skeletal model that simulatesthe movements in the original video. The differences are studied for the structural model and the prediction results are given for the original video. While others used special techniques represented in identifying each section of the body gradually, and this helps many people to get good accuracy, many researchers use a method based on these RGBD colors to color some parts of the body with different colors (Luvizon 2019), (Liu 2020) and use them to detect activity by color (Khowaja 2020). But the best way is to use the person as a shadow without any background or effects (Gnouma 2019); some other solutions are based on placing the skeleton in Space-Time and doing a 3D tracking (Yadav 2019). So, we can say this filed is missing two points: the first one, most of the old research is not in real-time, and the second one is the accuracy of the work according to all these results. It’s time to give our way to improve those problems, to get a real-time recognition with high accuracy.