Search

Multi-human tracking in the crowded environment is a challenging problem due to occlusions, pose change, viewpoint variation and cluttered background. In this work, we propose a robust feature learning for tracking-by-detection methods based on second-order attention network that can capture higher-order relationships between salient features at the early stages of Convolutional Neural Network (CNN). Guided Second-Order Attention Network (GSAN) that, unlike the existing attention learning methods which are weakly-supervised, uses a supervisory signal based on the quality of the self-learned attention maps. More specifically, GSAN looks into the attended maps of a person having the highest confidence and supervise itself to look into the correct regions in the images of the person. Attention maps learned this way are spatially aligned and thus robust to camera-view changes and body pose variations. We verify the effectiveness of our approach by comparing with the state-of-the-art methods on challenging person re-identification and multi object tracking (MOT) datasets.

Deep learning is the current state-of-the-art technique for most machine learning and data analytics tasks. It has been applied to all aspects of human life. Traditional methods are not competitive for solving pedestrian trajectory prediction problems due to the diversity of its environment and the uncertainty of the original trajectory. Due to the great success of the RNN (Recurrent Neural Networks) architecture in sequence prediction, it has become the first choice to us for solving pedestrian trajectory prediction problems in a crowded environment. In this work, in order to solve the problem that RNN architecture does not have great accuracy when the input is long sequence, we use LSTM (Long Short Term Memory) which produces higher accuracy when the input is long sequence. In the process of building our LSTM based prediction model, we find the best loss function through experiments and data analysis. Then we try various model hyperparameters combinations and find best hyperparameter values and parameter ranges that would make the prediction results more accurate. At the same time, we creatively use velocity values of pedestrian trajectory rather than coordinate values as input of the model. Experiments on several datasets shows that the proposed approach achieves high accuracy when it predicts pedestrian trajectory in a crowded environment.

Pedestrian motion are inherently multi-modal in nature influenced by presence of other human and physical objects in the environment. Traiectory prediction models need to address both human-human and human-space interaction issues. In this work, we leverage both pedestrians information and scene information of the navigation environment for jointly predicting trajectories of the pedestrian. We introduce a new Recurrent Neural Network based sequence model with attention mechanisms that address both human-human and human-space interaction challenges. The encoder encodes all the pedestrian trajectories and create a social context. The scene information of navigation environment is extracted using CNN and serves as a physical context for the model. Our approach utilizes physical and social attention mechanism to find semantic alignments between encoder and decoder. The social attention mechanism allow the model to look into similar step of pedestrian trajectory. The physical attention mechanism tells the model where and what to focus on the scene. Experiment on several datasets shows that the proposed approach which combine social and physical attention performs better than when this information is utilized independently.