Aiming at the problem that the object detector relies too much on the classification effect of convolutional network and cannot use motion information when detecting falls, this paper designs a fall detection model based on YOLOv5s and improved centroid tracking. To solve the problem of resource consumption, the MobileNetV3 network and Slim Neck module are used to lightweight YOLOv5s, and the SE module in the MobileNetV3 network is replaced with the more efficient ECA module, which reduces the network complexity while maintaining high accuracy. Hash sensing algorithm is introduced to improve centroid tracking, increase the basis of target association, and improve the accuracy of fall detection. The experimental results show that the size of the improved YOLOv5s model is reduced by 52.2%, the computational capacity is reduced by 51.8%, and the accuracy is as high as 90.3%. The accuracy of fall detection model with improved centroid tracking was increased by 4.3%. The results show the effectiveness and superiority of the proposed model.