Human Fall Detection Algorithm Based on Mixed Attention Mechanism

The health of older people is always a matter of concern, and falls can cause injuries and even death in severe cases. At present, the fall detection algorithm based on computer vision has a large amount of computation and is easy to be interfered by shielding objects. In addition, there are also problems such as a high rate of missed detection and poor real-time performance. This paper proposes a fusion hybrid attention mechanism of human body fall detection algorithm, a YOLOv3-tiny as a benchmark algorithm, add channel attention in the process of feature extraction and spatial attention mechanism, using the channel the different characteristics of attention to change the network the attention of the weight, the spatial attention change characteristics of pixels of attention weights in the figure, by increasing attention in the process of the fall detection selectivity of focusing not obscured human body parts, filter other essential features. Experimental results show that the optimized human fall detection algorithm improves the mAP of the test set by 8.93% compared with the benchmark YOLOv3-tiny algorithm and can also solve the problem of obstructing the human body with obstacles. In addition, this algorithm can detect falling and fall to the ground in real-time. On average, the falling action is detected 625.58ms in advance, and the FPS is 45.35. Therefore, it has practical feasibility and effectiveness.