13
Learn from IoT: Pedestrian Detection and Intention Prediction for Autonomous Driving Gürkan Solmaz, Everton Luis Berz, Marzieh Farahani Dolatabadi*, Samet Aytac, Jonathan Fürst, Bin Cheng, Jos den Ouden* IoT Research Group NEC Laboratories Europe * Eindhoven University of Technology (TU/e) 1st ACM Workshop on Emerging Smart Technologies and Infrastructures for Smart Mobility and Sustainability (SMAS), ACM MobiCom 2019 Los Cabos, Mexico, October 21, 2019 This work has received funding from the European Union’s Horizon 2020 research and innovation programme within the project AUTOPILOT (Automated Driving Progressed by Internet Of Things) under the grant agreement No 731993. Responsibility for the information and views set out in this document lies entirely with the authors.

Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

Learn from IoT: Pedestrian Detection and Intention Prediction for Autonomous Driving

Gürkan Solmaz, Everton Luis Berz, Marzieh Farahani Dolatabadi*, Samet Aytac, Jonathan Fürst, Bin Cheng, Jos den Ouden*

IoT Research Group

NEC Laboratories Europe

* Eindhoven University of Technology (TU/e)

1st ACM Workshop on Emerging Smart Technologies and Infrastructures for Smart Mobility and Sustainability (SMAS), ACM MobiCom 2019

Los Cabos, Mexico, October 21, 2019

This work has received funding from the European Union’s Horizon 2020 research and innovation programme within the project AUTOPILOT (Automated Driving Progressed by Internet Of Things) under the grant agreement No 731993. Responsibility for the information and views set out in this document lies entirely with the authors.

Page 2: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

2 © NEC Laboratories Europe GmbH 2019

Motivation

▌Why “learn from IoT”?Uber accident (2018): Both car sensors

and driver can easily fail

Mobile sensing, vast amount of sensors, RSUs, V2V, V2X communication, 5G, edge computing

IoT can support traffic safety

▌Major advancements in ML due to increased availability of data, computation, and software TensorFlow, PyTorch

▌IoT expansion to many domainsSmart cities and smart mobility

IoT platforms: Azure, AWS IoT

Large-scale tesbeds

▌Aim: Combine IoT and ML for enhanced AD safety

Page 3: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

3 © NEC Laboratories Europe GmbH 2019

Proposed approach

▌ System learning from data collected from

1) in-vehicle sensors (e.g., cameras)

2) external IoT sources such as mobile devices of VRUs

▌ Components included in the system setup:

1) autonomous car

2) mobile devices with the mobility app

3) in-vehicle IoT platform

4) VRU detection and intention prediction components

5) cloud IoT platform

▌ ML leverages data sources for training

Data collected by the vehicle at the TU/e campus

Smartphone apps of TU/e and NEC for pedestrian movement tracking

Mobile ITS-G5 location devices

▌ We propose methods for accurate VRU detection and pedestrian intention prediction

1) creating a world model (WM) by combining the vehicle and mobile device data and performing VRU detection and localization

2) using the latest trajectory data (in-vehicle or mobile device) to predict the pedestrians’ intended movement steps based on historical measurements

▌ Petri Net models to combine VRU detection and intention prediction

Decide under which conditions which inputs can be combined for the autonomous decision-making

Page 4: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

4 © NEC Laboratories Europe GmbH 2019

VRU detection through WM

▌WM contains the vehicle itself and objects around

▌The formalism adopted is WIRE where the

WM aims to track semantic objects such as VRUs http://wiki.ros.org/wire

(World Information for Robot Environments framework by TU/e)

▌Multiple Hypothesis Tracker (MHT) algorithm [2] combines evidences to a common world representation dynamically

▌Objects’ attributes, classification, and prior knowledge are associated in the hypotheses treeEvery hypothesis contains a list of anchors and has a correctness probability

Each anchor contains an individual symbol, a set of measurements and a probabilistic signature that consists out of a mixture of probability density functions generated by a set of behavior models

The predicate attribute space represents predicate grounding relations that link attribute values and predicate symbols [2]

Page 5: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

5 © NEC Laboratories Europe GmbH 2019

Pedestrian intention prediction method

▌This method predicts the next location of pedestrians based on historical data and the current position

▌ML model uses the mobile device data

Speed of pedestrian

GPS trajectory values

▌The model predicts nf future locations

▌The representation of pedestrian trajectory is inspired by the model in [8]

the path modeling with an

example pedestrian

trajectory during the TU/e

experiments

Page 6: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

6 © NEC Laboratories Europe GmbH 2019

Intention prediction ML model

▌ Three input features, three concatenation layers, and nf output features

▌ Adam algorithm [5] for the optimization process with ReLU activation function

▌ Embed encoder: Map the inputs into vectors and then forward to the (intermediary) concatenation layers

▌ Intermediary layers: Concatenate all outputs of the feature encoders and pass the concatenation through fully connected layers

▌ The data is randomly partitioned into training (70%), validation (20%), and testing (10%) subsets.

We defined np = 10 and nf = 5

Approximately equivalent to 11m

and 5.5m in a straight-walking

distance

Page 7: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

7 © NEC Laboratories Europe GmbH 2019

Combining VRU detection and intention prediction

▌ Combining the inputs using stochastic priority Petri Nets from the two previous parts to support the autonomous decision-making

States/places: big circles, transitions: rectangles

Probabilistic var. for transitions: λ

Priorities: Curly-braced numbers

# tokens: Numbers w/o curly braces

▌ The model has two types of VRUs

1. users of mobile devices and our app (w/ IoT data)

2. people w/o the app

▌ Three possible cases exist during the objects come to vicinity of each other

1) both the mobile device and vehicle sensor data,

2) only mobile device data

3) only vehicle sensor (camera) data available in the in-vehicle platform

Page 8: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

8 © NEC Laboratories Europe GmbH 2019

TU/e experiments

▌ We conduct the pilot tests mainly at the TU/e campus with 2km road network, speed limit 15km/h

▌ Three pilot tests: Each ~1-2 weeks long 21 controlled experiments (total of 70 runs)

Predefined pedestrian movements

Autonomous driving behaviors are mostly uncontrolled

▌ Custom-built autonomous car prototype (Toyota Prius)

▌ The car has an ITS-G5 device connected to the in-vehicle IoT platform Two pedestrians also carry mobile ITS-G5 devices

▌ IoT gateway connects the in-vehicle platform with two cloud IoT plaftormsvia MQTT and HTTP using cellular 4G

▌ Robot operating system (ROS) operates on the in-vehicle IoT platform

Page 9: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

9 © NEC Laboratories Europe GmbH 2019

Message delay and WM

▌Smartphone to vehicle delay ~0.6secFrom smartphone to cloud and lastly in-vehicle IoT platform (ROS timestamps)

▌In the WM, vehicle and mobile devices together has more consistency compared to only vehicleVehicle receives the global location of a pedestrian a few seconds before the

camera detection

Page 10: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

10 © NEC Laboratories Europe GmbH 2019

Intention detection results

▌ Ludwig and TensorFlow to train and validate the model

▌ Each step is a consecutive data point in the range of [0.5,1] sec Most predictions for 1st step is <1m

▌ Controlled scenarios: Walking-straight has the best accuracy Max error of 1.9m in the furthest prediction step

▌ 18% and 48% decrease in the error compared to the NARX and Dead Reckoning approaches 0.2m better on average than [8]

Page 11: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

11 © NEC Laboratories Europe GmbH 2019

Conclusion & future work

▌We propose learning from IoT data to improve the safety of autonomous driving

Key problems: pedestrian detection and pedestrian intention prediction

▌The proposed system can complement existing safety solutions

▌Future work:

extracting features from other data sources such as OpenStreetMap and traffic operation center data (e.g., traffic lights)

Old problems also still exist

Page 12: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform

12 © NEC Laboratories Europe GmbH 2019

This work has received funding from the European Union’s Horizon 2020 research and innovation programme within the project AUTOPILOT (Automated Driving Progressed by Internet Of Things) under the grant agreement No 731993. Responsibility for the information and views set out in this document lies entirely with the authors.

Questions?

[email protected]

Page 13: Learn from IoT: Pedestrian Detection and Intention ...gsolmaz/publications/MobiCom19... · Smartphone to vehicle delay ~0.6sec From smartphone to cloud and lastly in-vehicle IoT platform