Abstract: In recent years, it has become increasingly important to automatically detect falls and provide feedback in emergency situations. In complex scenes, aiming at the problem of the degradation of detection performance due to changes in illumination, occlusion and scale changes, a real-time and robust fall detection algorithm is proposed. Firstly, the YOLO v3 target detection module is used to complete pedestrian detection. Then, after extracting the deep features of each tracked bounding box in the tracking module, data enhancement and re-detection techniques are used to improve the detection accuracy under light changes, and the attention mechanism subnetwork is introduced to deal with the detection of obscured targets. Finally, the final fall judgment module is used to judge the pedestrian posture and complete real-time fall detection and alarm. The experimental results on the Cityperson data set, Montreal fall data set and self-built data set show that the detection accuracy of the pedestrian detection algorithm reaches 87.05%, the detection accuracy of the fall algorithm reaches 98.55%, and the delay is within 120 ms. Good performance can still be obtained under the influence of occlusion.

Key words: pedestrian detection, fall detection, deep learning, convolutional neural network, attention mechanism