Abstract
We describe and test an obstacle-detection system for small, lake-deployed autonomous surface vehicles (ASVs) that relies on a low-cost, consumer-grade camera and runs on a single-board computer. A key feature of lakes that must be accounted for is the frequent presence of the shoreline in images as well as the land-sky boundary. These particularities, along with variable weather conditions, result in a wide range of scene variations, including the possible presence of glint. The implemented algorithm is based on two main steps. First, possible obstacles are detected using an innovative gradient-based image processing algorithm developed especially for a camera with a low viewing angle to the water (i.e., the situation for a small ASV). Then, true and false positives are differentiated using correlation-based multi-frame analysis. The algorithm was tested extensively on a small ASV deployed in Lake Geneva. Under operational conditions, the algorithm processed 640×480-pixel images from a Raspberry Pi Camera at about 3—4 Hz on a Raspberry Pi 3 Model B computer. The present algorithm demonstrates that single-board computers can be used for effective and low-cost obstacle detection systems for ASVs operating in variable lake conditions.
Highlights
Autonomous Surface Vehicles (ASVs) are increasingly deployed in different applications, including for acquisition of field data and/or for monitoring inland water bodies, in particular lakes
We present an image-based obstacle detection algorithm for small ASVs deployed on lakes
Water patterns could more be detected as obstacles, the parameters controlling the update of likelihood of possible obstacle blobs (POBs) and obstacle blobs (OBs) must be tuned
Summary
Autonomous Surface Vehicles (ASVs) are increasingly deployed in different applications, including for acquisition of field data and/or for monitoring inland water bodies, in particular lakes. This map is multiplied element-wise with the gradient magnitude frame (f), resulting in Fig 3j, which enhances the gradient surrounding the previously found pixels (g), and allows for detection of horizontal edges Another threshold using a percentile of 1.5% segmented pixels (and a minimum threshold value of 0.05) is applied (k), after which a median filter is used to smooth the result, and a 4-neighbors flood filling algorithm is applied (l) to recover blobs for the following step. Relevant data for detected obstacles (angular position relative to the ASV heading, estimated distance based on the ASV attitude, etc.) are available for ASV navigation
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have