• ai
• alarm generator
• cnn
• computer vision
• deep learning
• embedded
• fpga
• neural networks
• nvidia jetson
• object detection
• raspberry
• real-time
• video surveillance
• xilinx

The aim of this work is to design and develop a neural network for video surveiilance applications, in particular smoke and fire detections. In particular, two neural networks have been developed: the first one is a regressor which has the task of identifying in the image space the ROIs that contain the possible anomalies while the second one is a classifier that has the task of deciding whether the anomaly is true or not. In addition there is an alarm generator algorithm based on the persistence of the anomaly over time. This architecture was then successfully implemented on several embedded systems like Raspberry Pi 3 and 4, NVIDIA Jetson nano and AGX Xavier and Xilinx ZCU104. A performance analysis was performed on each platform and in general the algorithm was more robust and efficient than the current state of the art considering both neural networks based techniques and classical techniques.