Using AI-based Edge Processing in Monitoring the Pedestrian Crossing

In edge processing, data collection devices are used to pre-filter data. As a result, only data of interest will be written to memory. This approach significantly reduces the amount of transferred data. Various algorithms, such as background segmentation or artificial intelligence (AI) techniques based on various neural networks, can be used to detect data of interest. This paper uses an AI-based technique in the edge processing environment to perform the learning process with a chosen neural network. The environment containing Nvidia Jetson Xavier NX is employed to train the MobileNetV3 network dedicated to detecting objects of interest like people or vehicles while monitoring the pedestrian crossing. The network consists of the initial fully connected convolution layer with 32 filters, followed by 19 residual bottleneck layers. This paper also proposes a learning process that uses the collected data after manual validation and significantly improves accuracy over the original network.