Han Cui

Applications for millimeter-wave (mmWave) radars have become increasingly popular in human activity recognition. Many researchers have combined radars with neural networks and gained a high performance on various applications. However, most of these studies feed the raw point cloud data directly into the networks, which can be unstable and inaccurate under certain circumstances. In this paper, we define a reliability measure of the point cloud data and design a novel voxelization algorithm to reconstruct the data. Experiments show that our algorithm can improve the stability of the point cloud generated from mmWave radars in terms of error reduction and scene re-construction. We demonstrate the effectiveness of our proposed algorithm using a neural network-based system for identifying a person’s sitting direction. In our experiment, compared with the baseline, our voxelization algorithm can improve the system in terms of accuracy (4.3%), training time (55.6%), and computational complexity, which is more suitable for light-weighted networks and low energy consumption platforms.

Millimetre-wave (mmWave) radar is increasing in popularity for human activity recognition, due to its advantages of high resolution, non-intrusive nature and suitability for various environments. In this paper, we present a novel human posture estimation system using mmWave radars. The system detects people with arbitrary postures in indoor environments at close distances (within two metres), and estimates the posture by localising the key joints. We use two mmWave radars to capture the scene and a neural network model to estimate the posture. The neural network model consists of a part detector that estimates the subject’s joint positions, and a spatial model that learns the correlation between the joints. A temporal correlation step is introduced to further refine the estimate when in real-time operation. The system can provide an accurate posture estimate of the person in real-time at 20 fps, with a mean localisation error of 12.2 cm and an average precision of 71.3%.

This paper studies a WiFi indoor localisation technique based on using a deep learning model and its transfer strategies. We take CSI packets collected via the WiFi standard channel sounding as the training dataset and verify the CNN model on the subsets collected in three experimental environments. We achieve a localisation accuracy of 46.55 cm in an ideal (6.5m × 2.5m) office with no obstacles, 58.30 cm in an office with obstacles, and 102.8 cm in a sports hall (40 × 35m). Then, we evaluate the transfer ability of the proposed model to different environments. The experimental results show that, for a trained localisation model, feature extraction layers can be directly transferred to other models and only the fully connected layers need to be retrained to achieve the same baseline accuracy with non-transferred base models. This can save 60% of the training parameters and reduce the training time by more than half. Finally, an ablation study of the training dataset shows that, in both office and sport hall scenarios, after reusing the feature extraction layers of the base model, only 55% of the training data is required to obtain the models’ accuracy similar to the base models.

In engineering education, most emphasis on teaching is centred around Content Knowledge and less on technology and pedagogy. In this paper, we present an e-learning system that enables students to provide real-time feedback during lectures and helps lecturers improve/adjust their teaching by knowing more about the students' needs. We also discuss the relevance of our system in pedagogy. Finally, a complete system prototype contributing to technology, pedagogy, and content knowledge will be demonstrated.

This paper investigates the potential of task-based parallelisation for application loops in the presence of cross-iteration dependencies. A novel task-based parallelisation scheme is implemented as an LLVM compiler pass that could automatically parallelise loops without programmer intervention. The parallelisation scheme is shown to be able to achieve significant performance improvements on computationally expensive applications with up to 61x speedup on 64 threads, and is capable of dealing with imbalanced workloads and arbitrary dependency patterns.

We present a real-time stereo vision system implemented on a low power-consumption Nvidia Jetson TX2 embedded platform with a GPU accelerator. We use a local matching method with Normalized Cross Correlation as the cost function, perform a set of optimisations to reduce computational cost and improve output quality. We use the KITTI dataset to evaluate our system. With image resolution of 1242x375, our system achieves an average of 155 FPS with 90 disparity levels on the TX2 platform.

Accurate calculation of a disparity map in real-time is a challenge in computer vision and autonomous applications due to the expensive computation cost. This paper describes a local temporal stereo method, using normalized cross correlation (NCC) as the matching cost function. To increase the performance in terms of speed without degrading the quality, a down sampling of the disparity map in the y-axis has been used, with a restricted search range along the y-axis. In addition to these, the search range is further restricted by referencing the disparity map from the previous frames. The algorithm is evaluated on a Texas Instruments (TI) KeyStone II embedded SoC with four ARM cores and eight DSP cores. The application has been written using OpenMP, OpenCL and IPC/MessageQ for parallelising the program. The core of the application has been optimised for this specific platform using linear assembly code in order to take advantage of SIMD instructions available with the DSP. The application achieved 77 FPS on a 1242×375 resolution video with 80 disparity levels using one ARM core and eight DSP cores.