I've been trying to do this for a while - take a timelapse from the front of a ferry near sunset. While I shoot on a tripod, the resulting sequence still needs a lot of stabilization because we are on a boat! This sequence started at 5:00pm and ended 2076 frames later at 6:43pm. Exposure went from f6.3 at 1/125th of a second to f6.3 at 1/2 of a second. The interval is 3 seconds because I learned from an earlier trip you need more intermediate frames if you want any chance to align between frames. The script to stabilize is in python and uses OpenCV and the SIFT algorithm to do the image alignment. I use a variety of alignment masks through the sequence so I can exclude things like the clouds and ocean but also other ships moving side to side. This script writes a transforms.trf that ffmpeg can use for stabilization. Back in Decemeber 2021, I took a ferry trip but hadn't yet figured out how to stabilize it. import cv2 import numpy as np import matplotlib.pyplot as plt import pprint import math f = open("transforms.trf", "w") f.write("VID.STAB 1 ") f.write("# accuracy = 15 ") f.write("# shakiness = 3 ") f.write("# stepsize = 4 ") f.write("# mincontrast = 0.200000 ") f.write("Frame 1 (List 0 []) ") img1 = cv2.imread('DSC_4877_FerryHg.jpg') mask = cv2.imread('mask2.png', cv2.IMREAD_GRAYSCALE) Mask_5182 = cv2.imread('Mask_5182.png', cv2.IMREAD_GRAYSCALE) mask_5725 = cv2.imread('mask_5725.png', cv2.IMREAD_GRAYSCALE) mask_5885 = cv2.imread('mask_5885.png', cv2.IMREAD_GRAYSCALE) mask_6670 = cv2.imread('mask_6670.png', cv2.IMREAD_GRAYSCALE) currentMask = mask img1 = cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY) #sift sift = cv2.SIFT_create() keypoints_1, descriptors_1 = sift.detectAndCompute(img1,mask) for n in range(2,2064): print("Frame " + str(n)) img2 = cv2.imread("DSC_{:04d}_FerryHg.jpg".format(4876 + n)) img2 = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY) if ( (4876 + n) == 5182 ): currentMask = Mask_5182 if ( (4876 + n) == 5725 ): currentMask = mask_5725 if ( (4876 + n) == 5885 ): currentMask = mask_5885 if ( (4876 + n) == 6670 ): currentMask = mask_6670 keypoints_2, descriptors_2 = sift.detectAndCompute(img2,currentMask) #feature matching bf = cv2.BFMatcher(cv2.NORM_L1, crossCheck=True) matches = bf.match(descriptors_1,descriptors_2) matches = sorted(matches, key = lambda x:x.distance) countPositions = 0 listOfPoints = "" for idx, val in enumerate(matches): distance = math.sqrt( (keypoints_2[val.trainIdx].pt[0]-keypoints_1[val.queryIdx].pt[0]) * (keypoints_2[val.trainIdx].pt[0]-keypoints_1[val.queryIdx].pt[0]) + ( keypoints_2[val.trainIdx].pt[1]-keypoints_1[val.queryIdx].pt[1]) * (keypoints_2[val.trainIdx].pt[1]-keypoints_1[val.queryIdx].pt[1] ) ) if ( val.distance > 299 and idx > 10 ): continue print(" " + str(idx) + " " + str(distance) + " " + str(val.distance) +" (" + str(keypoints_1[val.queryIdx].pt[0]) + "," + str(keypoints_1[val.queryIdx].pt[1]) + ") -> (" + str(keypoints_2[val.trainIdx].pt[0]) + "," + str(keypoints_2[val.trainIdx].pt[1]) + ")" ) if( countPositions > 0 ): listOfPoints = listOfPoints + "," # https://github.com/georgmartius/vid.stab/blob/master/src/serialize.c # if(fscanf(f,"(LM %hi %hi %hi %hi %hi %lf %lf", &lm.v.x,&lm.v.y,&lm.f.x,&lm.f.y,&lm.f.size, # &lm.contrast, &lm.match) != 7) { # LM = Local Motion # Sample: (LM 0 0 922 424 224 0.507735 0.263512) listOfPoints = listOfPoints + "(LM {:0.0f} {:0.0f} {:0.0f} {:0.0f} {:0.0f} {:5.3f} {:5.3f})".format( -(keypoints_2[val.trainIdx].pt[0]-keypoints_1[val.queryIdx].pt[0]), -(keypoints_2[val.trainIdx].pt[1]-keypoints_1[val.queryIdx].pt[1]), keypoints_1[val.queryIdx].pt[0], keypoints_1[val.queryIdx].pt[1], 48, (300.0 - val.distance) / 300.0 , 0.6 - (300.0 - val.distance) / 1000.0) countPositions = countPositions + 1 f.write("Frame {:d} (List {:d} [{}])".format(n,countPositions,listOfPoints)) img1 = img2 keypoints_1 = keypoints_2 descriptors_1 = descriptors_2 f.close()