Violent action classification in community-based surveillance is a particularly challenging concept in itself. The ambiguity of violence as a complex action can lead to the misclassification of violence-related crimes in detection models and the increased complexity of intelligent surveillance systems leading to greater costs in operations or cost of lives. This paper demonstrates a novel approach to performing automatic violence detection by considering violence as complex actions mitigating oversimplification or overgeneralization of detection models. The proposed work supports the notion that violence is a complex action and is classifiable through decomposition into more identifiable actions that could be easily recognized by human action recognition algorithms. A two-stage framework was designed to detect simple actions which are sub-concepts of violence in a two-stream action recognition architecture. Using a basic logistic regression layer, simple actions were further classified as complex actions for violence detection. Varying configurations of the work were tested, such as applying action silhouettes, varying activation caching sizes, and different pooling methods for post-classification smoothing. The framework was evaluated considering accuracy, recall, and operational speed considering its implications in community deployment. The experimental results show that the developed framework reaches 21 FPS operation speeds for real-time operations and 11 FPS for non-real-time operations. Using the proposed variable caching algorithm, median pooling results in accuracy reaching 83.08% and 80.50% for non-real-time and real-time operations. In comparison, applying max pooling results to recalls reached 89.55% and 84.93% for non-real-time and real-time operations, respectively. This paper shows that complex action decomposition is deemed to be an appropriate method through the comparable performance with existing efforts that have not considered violence as complex actions implying a new perspective for automatic violence detection in intelligent surveillance systems.

Traditional object counting systems use object detection methods to count objects. However, when objects are small, crowded, and dense, object detection may fail, leading to inaccuracies in counting. To address this issue, we propose a crowded object counting system based on density map estimation. While most density map estimation models employ encoder–decoder or multi-branch approaches to generate feature maps at different scales for obtaining an accurate density map, improving the accuracy of crowded object counting remains a challenge. In this paper, we propose a novel model that can generate more accurate density maps, utilizing the context-aware network as the primary structure and integrating the self-attention mechanism. There are three main contributions in this paper. Firstly, the self-attention mechanism is employed to improve the accuracy of density map estimation. Secondly, the missing vehicle labels in the TRANCOS database are relabeled, ensuring that the ground truth data are more complete than the original TRANCOS database, thus enabling the proposed novel model to have higher crowded object counting accuracy. Thirdly, the parameters of the self-attention mechanism are analyzed to obtain the optimum parameter combination. The experimental results demonstrate that the accuracy of crowded object counting can reach 85.9%, 90.0%, 83.4%, and 92.6% for the TRANCOS, relabeled TRANCOS, ShanghaiTech Part A, and Part B datasets, respectively. Furthermore, the ablation study for the context-aware network with self-attention mechanism analyzes the optimum parameter combination.