TraPHic: Trajectory Prediction in Dense and Heterogeneous Traffic Using Weighted Interactions

Abstract

We present a new algorithm for predicting the near-term trajectories of road agents in dense traffic videos. Our approach is designed for heterogeneous traffic, where the road agents may correspond to buses, cars, scooters, bi-cycles, or pedestrians. We model the interactions between different road agents using a novel LSTM-CNN hybrid network for trajectory prediction. In particular, we take into account heterogeneous interactions that implicitly account for the varying shapes, dynamics, and behaviors of different road agents. In addition, we model horizon-based interactions which are used to implicitly model the driving behavior of each road agent. We evaluate the performance of our prediction algorithm, TraPHic, on the standard datasets and also introduce a new dense, heterogeneous traffic dataset corresponding to urban Asian videos and agent trajectories. We outperform state-of-the-art methods on dense traffic datasets by 30%. Code for our implementation can be found on our project webpage.

Trajectory Prediction in Heterogeneous Environment via Attended Ecology Embedding

“…Moreover, the scene context in SDD covers a variety of outdoor places and contains more spatial constraints such as intersection and roundabouts on a university campus. There are other datasets of heterogeneous traffic such as Apolloscape [13] and TRAF [7], unfortunately, due to the lack of raw image and camera matrices, respectively, we could not make use of them in our paper. In addition to the heterogeneous dataset, we also evaluate the performance of AEE-GAN in the homogeneous environment on two commonly-used homogeneous datasets, i.e., ETH [27] and UCY [19] datasets.…”

“…However, the LSTMs fall short of capturing the spatial dependencies between the encoded features from each individual. Therefore, TraPHic [6] uses Convolutional Social Pooling LSTM [10] to learn the locally useful interactions. Nevertheless, when the number of interactions becomes large, e.g., prediction on homogeneous datasets (many pedestrians), the performance significantly degrades due to the unnecessary interaction modeling.…”

“…Existing endeavors have demonstrated the strong correlations between inferring the trajectories and the heterogeneous environments [6,8,23]. For example, TrafficPredict [23] assumes that the same kind of road-agents shares similar behavior characteristics; thus they introduce an LSTM-based 4D Graph, which an instance layer and a category layer are constructed to capture the similarity between different individuals.…”

“…For example, TrafficPredict [23] assumes that the same kind of road-agents shares similar behavior characteristics; thus they introduce an LSTM-based 4D Graph, which an instance layer and a category layer are constructed to capture the similarity between different individuals. TraPHic [6] models the different classes by taking into account the size of road-agents, and they introduce the convolutional LSTM model with the idea of horizon modeling, i.e., prioritizing the interactions in front of each road-agent. Nevertheless, the second and third challenges remain unsolved for the heterogeneous trajectory prediction task.…”

Data-driven travel time and demand estimation of vehicles in smart cities

“…In [46] [56], minimal on-road time route scheduling with parking facilities [57], as well as analysis of the public bus transportation systems [58]. In general, there is also a growing trend towards the designing of deep learning techniques to solve urban mobility problems, such as crowd-flow estimations [59], transportation mode prediction [60] and trajectory predictions for road agents [61][62][63][64][65].…”