Improved RNN-Maxout Framework for Location specific Human Trajectory Prediction with Yolov3 Tracking Model

Abstract

In applications such as autonomous vehicles, environmental planning and development, environmentally conscious automated machinery, and smart tracking systems, human trajectory forecasting is a crucial subject. Several issues surround this intricate topic, including multimodality, which human-human connection, human-space connection, and ability for generalization. Currently, modern facilities have not fully looked at these issues, which needs the advancement in forecasting model. By considering this, this paper presents an improved deep learning architecture for location specific human trajectory prediction along with YoloV3 based tracking model. At first, the input video is converted into a frames, and after that it is filtered to remove the noise via improved filter method as the pre-processing stage. Certain features are extracted from the pre-processed image (frame) including features like shape features, GLCM features, Improved median ternary pattern features and spatial temporal features. The extracted features are combined to determine the final features that are given as the input into the hybrid prediction model, which combines the models like Improved RNN and Deepmaxout model. Here the location is predicted from the hybrid model, and also to make the precise prediction, their weights are optimally tuned by SU-TSO method (Self Upgraded-Tuna Swarm Optimization). Finally, based on the location prediction, the tracking is carried out by the Yolo v3 model, which is an automatic location tracking system. Then the experimental investigation of the proposed work takes place, which shows its betterment than the other traditional methods with respect to error measures.