Artificial Intelligence for Analyzing Pedestrian Motion and Abnormal Situation by Thermal and RGB Camera

Zaixin Liu, Baoheng Zhang, Wilton W. T. Fok, Yingxian Chen

Abstract: To reduce the injury and fatality of the special group of pedestrians in the rapid traffic flow, engineers can utilize artificial intelligence to detect human bodies and postures. The design enables identification of the existence of pedestrians in the intersection and identifies their postures with NVIDIA CUDA Deep Neural Network (CuDNN). This mechanism helps the authority to devise an advanced algorithm to manipulate the traffic lights to allow these accommodation-needed people to cross the road in time when they are in the intersections. By experiments, Zaixin has found that the trend of altering a single parameter cannot be summarized without taking the influence of other parameters and the complexity of data sets into account. Enlarging the size of the image can increase the accuracy but can also decrease that in different scenarios. RGB models usually deliver more desirable results with higher maximum Average Precision (mAP) of human detection and the accuracy of the gesture recognition than the thermal model trained by our team. Trained models are more precise and accurate than the pre-trained models generated with HRNet. Implementations based on the new techniques can substantially enhance the safety of senior and disabled pedestrians.