In the past, I worked for many detection tasks using airborne images for the purpose of vision-based navigation of UAVs. One thing I found out is that when a vision system relies on the detection of an ‘entire’ target, the navigation is not really robust. From time to time during target chasing, there are moments that the vision system sees only the part of the target due to tracking errors and disturbances, etc., and that absolutely happens at some point of vision-based landing because the target gets too large to fit into an on-board image. One approach to solve this issue with traditional machine-learning-based algorithms such as Viola-Jones and Dalal Triggs frameworks would be to add occluded images to positive images, but a naive implementation would lose a lot of information. For example, for the purpose of vision-based landing, detecting an occluded target is not enough, and we need to know where the center of the target from occluded images. If we treat all the images that contain part of the target, we never know where the center of the target. My approach is to use the combination of object proposals and particle filter. We have many algorithms to evaluate ‘objectness’ such as selective search, BING, etc. Each proposed object is compared with the target image using HOG features, normalized cross correlation, SIFT to obtain the score of the each object, and the measurements are fused via a particle filter framework.