In recent years, robots and computer scientists have been working on various systems that can identify and traverse items in their environment. The majority of these systems rely on deep learning and machine learning algorithms that have been trained on big image datasets. There are enormous advantages of radars over optical sensors. Still, there are presently few image datasets for training machine learning models that comprise data obtained using radar sensors. Furthermore, many open-source radar datasets present today exhibit challenges when used for different user applications.
A team has recently created a new method for automatically generating datasets with tagged radar data-camera images. This method labels the radar point cloud using a highly accurate object recognition algorithm (YOLO) on the camera image stream and an association technique (the Hungarian algorithm). The method is likely to advance Deep Leaning Market as it would enable the technology to track objects with sensor fusion.
Deep-learning systems that use radar necessitate a large amount of labelled training data. However, labelling radar data is a time- and labour-intensive procedure that is often carried out by manually comparing it to a parallelly acquired image data stream. Suppose the camera and radar are staring at the same item. In that case, one may use an image-based object detection framework (YOLO) to automatically label the radar data instead of manually looking at images.
The team's technique's co-calibration, grouping, and association capabilities are three distinguishing qualities. The method co-calibrates a radar and its camera to figure out how the location of an object identified by the radar would be converted into digital pixels present on a camera.
In the future, this team of academics' novel approach could aid in the automated production of radar-camera and radar-only datasets. Furthermore, in their article, the researchers looked at both proof-of-concept classification techniques based on a radar-camera sensor-fusion approach and data acquired only by radars.
Instead of using simply the point-cloud distribution or just the micro-doppler data. The team suggested employing an effective 12-dimensional radar feature vector generated using a combination of spatial, Doppler, and RCS statistics.
Finally, the team stated that current research could open up new avenues for quick analysis and training of deep learning-based models for classifying or tracking objects utilising sensor-fusion. These models have the prospect to improve the performance of a wide range of robotic systems, from autonomous automobiles to small robots.