Narrow Results

Detecting collision-course targets in aerial scenes from purely passive optical images is challenging for a vision-based sense-and-avoid (SAA) system. Proposed herein is a processing pipeline for detecting and evaluating collision course targets from airborne imagery using machine vision techniques. The evaluation of eight feature detectors and three spatio-temporal visual cues is presented. Performance metrics for comparing feature detectors include the percentage of detected targets (PDT), percentage of false positives (POT) and the range at earliest detection ($R_{\mathrm{\text{det}}}$). Contrast and motion-based visual cues are evaluated against standard models and expected spatio-temporal behavior. The analysis is conducted on a multi-year database of captured imagery from actual airborne collision course flights flown at the National Research Council of Canada. Datasets from two different intruder aircraft, a Bell 206 rotor-craft and a Harvard Mark IV trainer fixed-wing aircraft, were compared for accuracy and robustness. Results indicate that the features from accelerated segment test (FAST) feature detector shows the most promise as it maximizes the range at earliest detection and minimizes false positives. Temporal trends from visual cues analyzed on the same datasets are indicative of collision-course behavior. Robustness of the cues was established across collision geometry, intruder aircraft types, illumination conditions, seasonal environmental variations and scene clutter.

In traditional conveyor belt edge detection methods, contact detection methods have a high cost. At the same time noncontact detection methods have low precision, and the methods based on the convolutional neural network are limited by the local operation features of the convolution operation itself, causing problems such as insufficient perception of long-distance and global information. In order to solve the above problems, a dual flow transformer network (DFTNet) integrating global and local information is proposed for belt edge detection. DFTNet could improve belt edge detection accuracy and suppress the interference of belt image noise. In this paper, the authors have merged the advantages of the traditional convolutional neural network’s ability to extract local features and the transformer structure’s ability to perceive global and long-distance information. Here, the fusion block is designed as a dual flow encoder–decoder structure, which could better integrate global context information and avoid the disadvantages of a transformer structure pretrained on large datasets. Besides, the structure of the fusion block is designed to be flexible and adjustable. After sufficient experiments on the conveyor belt dataset, the comparative results show that DFTNet can effectively balance accuracy and efficiency and has the best overall performance on belt edge detection tasks, outperforming full convolution methods. The processing image frame rate reaches 53.07 fps, which can meet the real-time requirements of the industry. At the same time, DFTNet can deal with belt edge detection problems in various scenarios, which gives it great practical value.