Predicting humans future motion trajectories in video streams using generative adversarial network

Understanding the behavior of human motion in social environments is important for various domains of a smart city, e.g, smart transportation, automatic navigation of service robots, efficient navigation of autonomous cars and surveillance systems. Examining past trajectories or environmental factors alone are not enough to address this problem. We propose a novel methodology to predict future motion trajectories of humans based on past attitude of individuals, crowd attitude and environmental context. Many researchers have proposed different techniques based on different features extraction and features fusion to predict the future motion trajectory. They used traditional machine learning algorithms like SVM,social forces, probabilistic models and LSTM to analyze the heuristic motion trajectories but they didn’t consider the other environmental factors e.g relative positions of other humans present in environment and positions of objects present in environment which can affect the motion trajectories of humans. We intend to achieve this goal by employing Long Short Term Memory(LSTM) units to analyze motion histories, convolution neural networks to environmental facts e.g. human-human, human-object interaction and relative positioning of 80 different objects including pedestrians and generative adversarial networks(GANs) to predict possible future motion paths. Our proposed method achieved 70% lower Average Displacement Error(ADE) and 41% lower Final Displacement Error(FDE) in comparison to other state of the art techniques.