Anomalous behavior detection is a challenging research area within computer vision. One such behavior is throwing action in traffic flow, which is one of the unique requirements of our Smart City project to enhance public safety. This paper proposes a solution for throwing action detection in surveillance videos using deep learning. At present, datasets for throwing actions are not publicly available. To address the use-case of our Smart City project, we first generate the novel public 'Throwing Action' dataset, consisting of 271 videos of throwing actions performed by traffic participants, such as pedestrians, bicyclists, and car drivers, and 130 normal videos without throwing actions. Second, we compare the performance of different feature extractors for our anomaly detection method on the UCF-Crime and Throwing-Action datasets. Finally, we improve the performance of the anomaly detection algorithm by applying the Adam optimizer instead of Adadelta, and we propose a mean normal loss function that yields better anomaly detection performance. The experimental results reach an area under the ROC curve of 86.10 for the Throwing-Action dataset, and 80.13 on the combined UCF-Crime+Throwing dataset, respectively.
Tire defect detection has significant industrial value and has been a research topic in both academia and industry. Despite its importance, prior works does not considered the practical manufacturing circumstances, where there are only limited annotation for the defect. Such limitation hinders the prior works from deploying to the real-world system. To address the problem of Tire Defect Detection with Limited Annotation (TTDLA), we proposed a novel framework, denoted as tire defect detection with Self-Supervision and Synthetic data (or S3). S3 first uses self-supervised learning to train the encoder without using any labeled data in the pretraining stage. The encoder is then adopted as the encoder of the Faster-RCNN detector in the fine-tuning stage. In addition, we proposed an algorithm to generate synthesized image by pasting defects randomly onto the regular image. Experiments demonstrate that both self-supervised learning and synthesized data boost the performance of the detector under TTDLA scenario.
Traffic signals are part of our critical infrastructure and protecting their integrity is a serious concern. Security flaws in traffic signal systems have been documented and effective detection of exploitation of these flaws remains a challenge. In this paper we present a visual analytics approach to look for anomalies in traffic signal data (i.e., abnormal traffic light patterns) that may indicate a compromise of the system. To our knowledge it is a first time a visual analytics approach is applied for the processing and exploration of traffic signal data. This system supports level-of-detail exploration with various visualization techniques. Data cleaning and a number of preprocessing techniques for the extraction of summary information (e.g., traffic signal cycles) of the data are also performed before the visualization and data exploration. Our system successfully reveals the errors in the input data that would be difficult to capture with simple plots alone. In addition, our system captures some abnormal signal patterns that may indicate intrusions into the system. In summary, this work offers a new and effective way to study attacks or intrusions to traffic signal control systems via the visual analysis of traffic light signal patterns.