Coupling Intent and Action for Pedestrian Crossing Behavior Prediction

Authors: Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, Xiaoxiao Du

IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated our approach on two naturalistic driving datasets, PIE and JAAD, and extensive experiments show significantly improved and more explainable results for both intent detection and action prediction over state-of-the-art approaches.
Researcher Affiliation Academia Yu Yao , Ella Atkins , Matthew Johnson-Roberson , Ram Vasudevan and Xiaoxiao Du University of Michigan {brianyao, ematkins, mattjr, ramv, xiaodu}@umich.edu
Pseudocode No The proposed architecture is illustrated in Figure 2.
Open Source Code Yes Our code is available at: https://github.com/umautobots/ pedestrian intent action detection
Open Datasets Yes We evaluate the effectiveness of our method for intent detection and action prediction on two publicly available naturalistic traffic video benchmarks, Pedestrian Intent Estimation (PIE) [Rasouli et al., 2019] and Joint Attention in Autonomous Driving (JAAD) [Kotseruba et al., 2016].
Dataset Splits Yes The PIE dataset was collected with an on-board camera covering six hours of driving footage. There are 1,842 pedestrians (880/243/719 for training/validation/test) with 2-D bounding boxes annotated at 30Hz with behavioral tags. Ego-vehicle velocity readings were provided using gyroscope readings from the camera. The JAAD dataset contains 346 videos with 686 pedestrians (188/32/126) captured from dashboard cameras, annotated at 30Hz with crossing intent annotations.
Hardware Specification Yes All models were trained with sample length T = 30, prediction horizon δ = 5, learning rate 1e 5, and batch size 128 on a NVIDIA Tesla V100 GPU.
Software Dependencies No The proposed neural network model was implemented using Py Torch.
Experiment Setup Yes All models were trained with sample length T = 30, prediction horizon δ = 5, learning rate 1e 5, and batch size 128 on a NVIDIA Tesla V100 GPU.