Date

2022-6-28

Author

Coşkun, Muhammet Furkan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

299
views

422
downloads

Cite This

Crowd analysis related tasks have a wide range of real-life applications such as mass surveillance, safety monitoring, disaster management, traffic control and public area design. There are different novel works on crowd counting, localization and abnormal event detection tasks in the computer vision literature. In this study, we propose a ConvLSTM based Convolutional Neural Network to solve these three problems together in a single network. Our proposed model is a combination of P2P-Net architecture, which is a crowd counting and localization architecture, with a novel crowd anomaly detection module. P2P-Net architecture uses a VGG-16 network as the backbone and directly predicts the point proposals that represent the human heads. The proposed crowd anomaly head consists of an LSTM encoder with two LSTM cells and a fully convolutional anomaly binary classifier head. The proposed architecture has single backbone, multiple head structure and makes predictions for three different crowd tasks: crowd counting, crowd localization and crowd anomaly detection. We train the whole model in an end-to-end manner using synthetic data. Since counting and localization in crowd problem requires pixel-level annotations, real-world dataset creation process involves significant manual effort. Moreover, since an unified definition for crowd anomaly situation has not existed yet, there are different crowd anomaly datasets that have different anomaly definitions. For this reason, multiple datasets couldn't be used together to train a network. Moreover, none of them are big enough to train a decent ConvLSTM network. To overcome this data problem, we use synthetic datasets for network training. There is a synthetic dataset that fits our problem, called GTA-Events dataset. However, since the amount of data and scene variations is not enough in this dataset, we created a realistic synthetic dataset using the GTA-V game. Our novel synthetic dataset, named as METU Synthetic Crowd Dataset(METU-SCD), includes different scenarios, weather conditions and time intervals of the day. We train our proposed model with our novel synthetic dataset along with the GTA-Events dataset. Results show that the proposed model reaches over 90\% accuracy in crowd anomaly detection task and outperform optical-flow based methods. In crowd counting and localization tasks, it outperform the base model(P2P-Net) on synthetic data, whereas the real data results are somewhat behind than other methods. We, also, conduct an ablation study on the effect of counting and regression losses to anomaly performance. It is shown that training the network with a combination of counting, regression and anomaly losses brings increase in anomaly prediction score compared to the case that only anomaly loss is used.

Subject Keywords

crowd analysis, multi task prediction, crowd counting and localization, convLSTM encoder based crowd anomaly prediction, crowd abnormal event detection, synthetic crowd dataset generation with GTA-V game, METU sytnhetic crowd dataset

URI

https://hdl.handle.net/11511/98176

Collections

Graduate School of Natural and Applied Sciences, Thesis