Drivers and pedestrians engage in non-verbal and social cues to signal their intent, which is crucial to their interactions in traffic scenarios. We propose to learn such cues and model a pedestrian’s intent. The learnt model is used to predict actions likely to be performed 400 - 600ms in the future. Responding to adverse actions in advance, we tread towards full autonomy.
Behavioral psychology characterizes every individual with a set of preferences. Groups exhibit emergent capabilities like that of schooling in fish or flocking in birds by integrating disparate preferences. This work is an investigation of swarm intelligence to evolve an autonomous vehicle convoying behaviour. Results demonstrate the consensus achieved by vehicle convoys to manoeuvre traffic lanes on highways. Convoys are shown to cruise at the maximum system speed, enhancing highway throughput and delivering optimal performance per vehicle.
3D Person Re-Identification
Investigation of deep learning networks to identify human subjects scanned by a Microsoft Kinect sensor. Person identification has numerous applications in robotics (identifying an owner) and surveillance (distinguishing allies from enemies). Security applications (such as identifying criminals from security footage) are of particular interest, and depth-based identification will work even if an individual covers their face. 3D sensors are affordable and can be used almost anywhere. The success of deep learning suggests that such networks would be adept at distinguishing between different individuals scanned by a 3D sensor. The project is implemented with 3D convolutional neural network as a feature extractor to identify people from the 3D pointcloud data. Pointcloud as voxelized 3D grids and as spatially encoded RGB images were both investigated on the Berkeley MHAD dataset.