Abnormal falls in public places have significant safety hazards and can easily lead to serious consequences, such as trampling by people. Vision-driven fall event detection has the huge advantage of being non-invasive. However, in actual scenes, the fall behavior is rich in diversity, resulting in strong instability in detection. Based on the study of the stability of human body dynamics, the article proposes a new model of human posture representation of fall behavior, called the “five-point inverted pendulum model”, and uses an improved two-branch multi-stage convolutional neural network (M-CNN) to extract and construct the inverted pendulum structure of human posture in real-world complex scenes. Furthermore, we consider the continuity of the fall event in time series, use multimedia analytics to observe the time series changes of human inverted pendulum structure, and construct a spatio-temporal evolution map of human posture movement. Finally, based on the integrated results of computer vision and multimedia analytics, we reveal the visual characteristics of the spatio-temporal evolution of human posture under the potentially unstable state, and explore two key features of human fall behavior: motion rotational energy and generalized force of motion. The experimental results in actual scenes show that the method has strong robustness, wide universality, and high detection accuracy.

This article proposes an automated machine learning method for semantic segmentation that can be used for automated training of models in fields such as autonomous driving. This method is not specific to a particular semantic segmentation model. Users can simply upload a dataset with the semantic segmentation model they use, and then choose to use this method. This method implements end-to-end machine learning from sensing data to semantic segmentation results and model evaluation. It integrates four main components: unsupervised data reduction through feature extraction and clustering, a contrastive learning-based evaluator of semantic segmentation results, interactive reinforcement learning-based data selection, and automatic hyperparameter tuning through Bayesian optimization. We demonstrate the practicality of this method on the MRSI and Cityscapes datasets, and trained mainstream semantic segmentation models, such as BiSeNet and STDC. Our results show that this method can effectively guide semantic segmentation training and reduce the training time by more than 20 $\%$ .

Point cloud densification is essential for understanding the 3D environment. It provides crucial structural and semantic information for downstream tasks such as 3D object detection and tracking. However, existing registration-based methods struggle with dynamic targets due to the incompleteness and deformation of point clouds. To address this challenge, we propose a Kalman-based scene flow estimation method for point cloud densification and 3D object detection in dynamic scenes. Our method effectively tackles the issue of localization errors in scene flow estimation and enhances the accuracy and precision of shape completion. Specifically, we introduce a Kalman filter to correct the dynamic target’s position while estimating long sequence scene flow. This approach helps eliminate the cumulative localization error during the scene flow estimation process. Extended experiments on the KITTI 3D tracking dataset demonstrate that our method significantly improves the performance of LiDAR-only detectors, achieving superior results compared to the baselines.

Dense crowd counting has become an essential technology for urban security management. The traditional crowd counting methods mainly apply to the scene with a single view and obvious features but cannot solve the problem with a large area and fuzzy crowd features. Therefore, this paper proposes a crowd counting method based on high and low view information fusion (HLIF) for large and complex scenes. First, a neural network based on an attention mechanism (AMNet) is established to obtain a global density map from a high view and crowd counts from a low view. Then, the temporal correlation and spatial complementarity between cameras are used to calibrate the overlap areas of the two images. Finally, the total number of people is calculated by combining the low‐view crowd counts and the high‐view density map. Compared to single‐view crowd counting methods, HLIF is experimentally more accurate and has been successfully applied in practice.

Vision-based track detection is an important part to realize driverless rail transportation, while the traditional track detection algorithm with manual feature extraction has drawbacks in both accuracy and speed, and the environmental impact factors such as light and water accumulation lead to the non-replication of the ground lane line detection algorithm. Aiming at the above problems, an improved bilateral segmentation deep learning network for real-time detection of rail transit is proposed. Firstly, the overall structure of the detection network is given. Secondly, the convolution layer is fused into the spatial path to make the segmentation and positioning of the track area more accurate. Then the semantic path is added to the end of the model to optimize the attention module, and the upsampling operation is omitted, so as to obtain a maximum field of sensitivity. Finally, the experimental verification is carried out on the given dataset. The result shows that the proposed network can effectively extract the track regions with a detection rate of 42.5 fps and a detection accuracy of 77.57 .

Quad-polarimetric SAR data has been proved to be useful for marine oil spill classification. Different SAR polarimetric features have been proposed to discriminate between oil spills and look-alikes which could cause false detection. In this paper we explored the ability of convolutional neural network (CNN) in automatic oil spill classification, by taking the advantage of H/A/Alpha polarimetric decomposition features and co-polarized correlation coefficients(CC). The convolutional neural network (CNN) was refined to realize the classification, in which global average pooling layer is applied instead of full connection layer. The quad-polarimetric Radarsat-2 data acquired during the Norwegian oil-on-water exercise was tested in the experiment. Sea surface was classified as clean sea, oil spill, look-alikes(biological oil spill in this case), and emulsion. The experiment results show that H/A/Alpha parameters and the combination of H/A/Alpha and co-polarized CC obtained higher accuracy, and the refined CNN has better performance than the traditional one in terms of accuracy and efficiency.