Socially Aware Trajectory Prediction For Multi Pedestrian Environment
| dc.contributor.author | Acharya, Bikram | |
| dc.date.accessioned | 2022-01-25T09:47:42Z | |
| dc.date.available | 2022-01-25T09:47:42Z | |
| dc.date.issued | 2021-08 | |
| dc.description | Human trajectory is the path that a human subject would most likely take to reach a speci c destination. | en_US |
| dc.description.abstract | Human trajectory is the path that a human subject would most likely take to reach a speci c destination. This route is estimated or predicted using pedestrian trajectory forecasting techniques. The key is to accurately encode observation sequence, model long-term dependencies from the past trajectories and forecast potential trajectories. Such models helps to learn social impact from other pedestrians, scene limits, and multi-modal possibilities of expected routes and can generalize to challenging scenarios and even output unacceptable solutions. In this thesis , a novel approach of e ective hard sampling with contrastive learning to preserve motion representation, which captures desirable generalization properties with little computational overhead is achieved. Further, improving the quality of visual representations in socially aware pedestrian trajectory prediction. ETH-UCY a benchmark dataset, comprising of total 5 di erent sets ETH, Hotel, Univ, Zara1 and Zara2 and TrajNet++, another benchmark consisting ETH, UCY, WildTrack, LCAS, and CFF dataset, are used for this thesis. Average Displacement Error (ADE), Final Displacement error (FDE) and Collision Avoidance Metric (CAM) are metrics used for performance evaluation. Experiments were carried out using real-world data and compared to state-of-art to assess the quality of the forecasting algorithm and the e ectiveness of process. The result shows that proposed methododology with hard sampling has better collision avoidance in 3 of the 5 sets of ETH-UCY dataset with collision values of Hotel(0.07), Univ(2.62) and Zara1(0.04) compared to that of existing social-nce model Hotel(0.38), Univ(3.08) and Zara1(0.18). Similarly, for TrajNet++'s the result shows better collision avoidance with CAM values Directional-LSTM(4.42) and Social-LSTM(5.20) in comparison to state-ofart. The obtained results indicates a considerable improvement in the accuracy of trajectory predictions with better collison avoidance. | en_US |
| dc.identifier.citation | MASTER OF SCIENCE IN INFORMATION AND COMMUNICATION ENGINEERING | en_US |
| dc.identifier.uri | https://hdl.handle.net/20.500.14540/7670 | |
| dc.language.iso | en | en_US |
| dc.publisher | Pulchowk Campus | en_US |
| dc.subject | Hard Negative Sampling, | en_US |
| dc.subject | Pedestrian Trajectory, | en_US |
| dc.subject | Contrastive Learning, | en_US |
| dc.subject | Motion Representation, | en_US |
| dc.subject | Multi Pedestrian Environment, | en_US |
| dc.subject | Trajectron++, | en_US |
| dc.subject | TrajNet++ | en_US |
| dc.title | Socially Aware Trajectory Prediction For Multi Pedestrian Environment | en_US |
| dc.type | Thesis | en_US |
| local.academic.level | Masters | en_US |
| local.affiliatedinstitute.title | Pulchowk Campus | en_US |
| local.institute.title | Institute of Engineering | en_US |