Md. Tareq Mahmood

In this paper, we propose a novel data-driven approach for a trip planner, that finds the most popular multi-modal trip using public transport from historical trips, given a source, a destination, and user-defined constraints such as time, minimum switches, or preferred modes of transport. To solve the most popular trip and its variants, we propose a multi-stage deep learning architecture, PathOracle, that consists of two major components: KSNet to generate key stops, and MPTNet to generate popular path trips from a source to a destination passing through the key stops. We also introduce a unique representation of stops using Stop2Vec that considers both the neighborhood and trip popularity between stops to facilitate accurate path planning. We present an extensive experimental study with a large real-world public transport based commuting Myki dataset of Melbourne city, and demonstrate the effectiveness of our proposed approaches.

Reconstructing a layout of indoor spaces has been a crucial part of growing indoor location based services. One of the key challenges in the proliferation of indoor location based services is the unavailability of indoor spatial maps due to the complex nature of capturing an indoor space model (e.g., floor plan) of an existing building. In this paper, we propose a system to automatically generate floor plans that can recognize rooms from the point-clouds obtained through smartphones like Google's Tango. In particular, we propose two approaches - a Recurrent Neural Network based approach using Pointer Network and a Convolutional Neural Network based approach using Mask-RCNN to identify rooms (and thereby floor plans) from point-clouds. Experimental results on different datasets demonstrate approximately 0.80-0.90 Intersection-over-Union scores, which show that our models can effectively identify the rooms and regenerate the shapes of the rooms in heterogeneous environment.

Trajectory User Linking (TUL) is the problem of identifying the user (i.e., his identity) from the trajectories generated by him. Existing works on TUL leverage a single trajectory for identifying a user. We propose a novel problem called Multi-Trajectory User Linking (MTUL), which leverages all available trajectories generated by a particular user to identify him. Thus, MTUL is essentially the generalized TUL problem. This problem has significant applications in Location-Based Services (LBSs) such as personalized route planning and point-of-interests (POI) recommendation, movement anomaly detection, etc. We provide an end-to-end solution to the MTUL problem using sequence embedding and GRU and achieve reasonable accuracy by taking into account the POI type and region information. We consider this work to be an important addition to the TUL research.

Handwritten digit recognition has been the “hello world” of deep learning. Yet, there are no significant work on Bengali handwritten digits due to a lack of benchmark dataset. NumtaDB is the largest dataset on Bengali handwritten digits and currently we have the best accuracy of 99.3359% on it. We used popular CNN architectures namely, ResNet34 and Resnet50. We preprocessed the data, used data augmentation, and trained our models with augmented data. We tested our models on both the raw test data and cleaned test data. We found that slightly underfitted models work better on the test data. And finally ensembled our six best models to get our final predictions. In this paper we describe some methods and techniques that performs well in NumtaDB dataset.