Object detection in surveillance system has to meet two requirements: high accuracy and realtime response. To achieve high accuracy, we adopt TensorBox architecture [1] as the realization that integrates the Long Short Time Memory (LSTM) of Recurrent Neural Network (RNN) into Convolutional Neural Network (CNN). Such the deep neural networks have been already known for their superior ability of object detection. Besides, the deep neural networks are usually computation-intensive and demand more resources to meet the realtime requirement. However, the system configuration is not properly selected and often leads to low utilization of the resources, which in turn makes the redundancy in the system with the expensive cost. In our performance investigation, we use the drone to capture the video and propose it as the scenario of a surveillance system. In addition, we experiment on different number of CPU cores and GPU with different available memory. Our performance investigation concludes that GPU outperforms CPU even though we use all of the CPU cores and the parameter tuning doesn't matter much with respect to the time.
展开▼
