Embedded Feature-Line Points Tracker for real-time Visual Odometry-based system

Line segment tracker is a fundamental technique in computer vision that follows linear structures in man-made environments, playing a crucial role in Visual Odometry (VO) applications. Although feature-line segments are more efficient than feature-corners in image understanding and geometric inference by enhancing orientation change cues, they face critical challenges such as over-segmentation, occlusions, and ambiguities. In this paper, we address line tracking issues by leveraging feature-line properties. We propose an efficient sparse Lucas-Kanade Optical Flow (OF)-based method to track line segment points. By utilizing the line-fitting model, the anchor, and the gradient properties, we enhance optical flow predictions through three methods: i) an enhanced RANSAC method that uses the segment anchor to reject outliers, ii) a minimization technique that adjusts OF predictions based on anchor direction and gradient, and iii) a flow correction method that integrates both RANSAC and minimization approaches. To meet real-time and embedded constraints, we adopt the Algorithm Architecture Adequacy (A3) methodology for suitable implementation in GPU-based heterogeneous architecture. We also conduct qualitative and quantitative benchmarks in indoor/outdoor environments against state-of-the-art VO front-ends. The results demonstrate the effectiveness of our proposed methods, highlighting the efficiency of tracking line segments via the anchor in real-time at about 50 fps within GPU-based embedded heterogeneous architectures.