You Are How You Drive: Peer and Temporal-Aware Representation Learning for Driving Behavior Analysis Pengyang Wang, Yanjie Fu, Jiawei Zhang, Pengfei Wang, Yu Zheng, Charu Aggarwal
Outline 1 ¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion
Background and Motivation 2 ¨ Car accident facts
Driving Behaviors 3
Driving Behaviors 4 It is essential to learn the pattern of driving behaviors
Challenges & Insights 4 ¨ Challenge I: GPS traces – Non-applicable GPS traces (e.g., time, latitude, longitude) encode the driving operations, states, and styles in a semantically implicit way ¨ Insight I: Transforming GPS traces into graphs Convenient for representation learning
Challenges & Insights 7 ¨ Challenge II: How to model dependencies? peer dependencies temporal dependencies ¨ Insight II jointly model the graph-graph peer dependency across drivers, as well as the current-past temporal dependency within a driver, in representation learning.
Outline 8 ¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion
Definition I 9 ¨ Driving Operation Driving operations are defined as a set of activities and steps that a driver operates when driving a vehicle, according to the driver’s personal judgment, experience and skills. Speed-related: acceleration, deceleration, constant speed Direction-related: turning left, turning right, moving straight
Definition II 10 ¨ Driving State A driving state concerns the way that a vehicle moves at a specific time point or in a small time window. In other words, a driving state of a vehicle contains both the speed status (i.e., acceleration, deceleration, constant speed) and the direction status (i.e., turning left, turning right, moving straight) of a vehicle. For instance, a driving state example of a car can be <constant speed, moving straight>.
Definition III 11 ¨ Driving State Transition Graph Driving State 1 Driving State 2 Driving State 5 Driving State 3 Driving State 4
Problem Statement 11 ¨ Given o a driver (a vehicle) o corresponding GPS trajectories D = [ < t, ϕ t , λ t > ] T t =1 ¨ Objective o learning a mapping function f : D → V a sequence of time-varying yet V = [ v n ] N n =1 relational vectorized representations ¨ Core tasks o Constructing multi-view driving state transition graphs o Automated profiling of driving behavior via peer and temporal-aware representation learning o Applications to transportation safety
Framework Overview 12 t=1 t=2 t=3 t=T � � Peer Dependency Transition Probability View Prediction and Historical Encoder Decoder …… Assessment of Driving Scores Temporal � Representation Result Dependency Transition Duration View (Vectors) GPS Trajectory Driving State Transition Graph Sequence Risky Area Detection PTARL
Outline 13 ¨ Background and Motivation ¨ Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion
Methodology 14 ¨ Construction of multi-view driving state transition graphs ¨ Peer and temporal-aware representationlearning
Construction of multi-view driving state transition graphs 15 v Detecting Driving Operations v Detection of driving-related operations. v Detection of direction-related operations. v Extracting Driving State Sequences (1)acceleration while turning right, (2)acceleration while turning left, (3)acceleration while straightforward, (4)deceleration while turning right, (5)deceleration while turning left, (6)deceleration while straightforward, (7)constant speed while turning right, (8)constant speed while turning left, (9)constant speed while straightforward
Construction of multi-view driving state transition graphs 17 v Constructing Multi-view Driving State Transition Graphs Driving State 1 Driving State 1 Driving State 2 Driving State 2 Driving State 5 Driving State 5 Driving State 3 Driving State 4 Driving State 3 Driving State 4 Transition probability view Transition duration view
Peer and temporal-aware representation learning 18 ¨ Intuition 1: Structural Reservation ¨ Intuition 2: Temporal Dependency ¨ Intuition 3: Peer Dependency
For Intuition 1: Structural Reservation 19 ¨ Base Model - Autoencoder = σ ( ˆ W o +1 z i + ˆ y 1 = σ ( W 1 x i + b 1 ) , y o b o +1 ) , ˆ i i = σ ( ˆ i + ˆ = σ ( W k y k − 1 y k + b k ) , ∀ k ∈ { 2 , 3 , · · · , o } , y k − 1 W k ˆ y k b k ) , ∀ k ∈ { 2 , 3 , · · · , o } , ˆ i i i = σ ( ˆ i + ˆ = σ ( W o +1 y o i + b o +1 ) . W 1 ˆ y 1 b 1 ) . ˆ z i x i y 1 y 1 ˆ ˆ x z x d 1 d 2 … … … … … … … d N input matrix D
For Intuition 2: TemporalDependency 20 # Sequential # Sequential Encode Step Decode Step = σ ( ˆ i + ˆ ( y 1 = σ ( W 1 x τ i + b 1 ) , y o W o +1 z τ b o +1 ) , i ) τ (ˆ i ) τ ) τ + b k ) , ∀ k ∈ { 2 , 3 , · · · , o } , i ) τ + ˆ = σ ( ˆ = σ ( W k ( y k − 1 y k − 1 ( y k W k (ˆ y k b k ) , ∀ k ∈ { 2 , 3 , · · · , o } , i ) τ (ˆ ) τ i i i ) τ + ˆ = σ ( ˆ = (1 − c τ ) z τ − 1 W 1 (ˆ y 1 b 1 ) . + c τ ˜ ˆ z τ z τ i . x τ i i i
For Intuition 3: Peer Dependency 21 2 X X H c ( G τ ) = s τ i,j · k z τ i � z τ j k 2 u i 2 U u j 2 U ,u i 6 = u j the similarity of driving behavior between the driver 𝒗 𝒋 and 𝒗 𝒌 at the time slot 𝝊 .
Objective Function 22 min 1 2 X X i � ˆ { k ( x τ i ) k 2 + α · H c ( G τ ) } x τ 2 τ ∈ T u i ∈ U ( n ) Temporal Dependencies Peer Dependencies
Outline 13 ¨ Background and Motivation ¨ Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion
Application 24 ¨ Prediction and Historical Assessment of Driving Scores ¨ Risky Area Detection
Outline 25 ¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion and Future Work
Evaluation 26 ¨ Data Description From Beijing City Driving Score Distribution
Evaluation 27 ¨ Baselines (1) Auto-Encoder: minimizes the loss between the original feature representations and reconstructed ones. (2) DeepWalk: uses local information obtained from truncated random walks to learn latent representations. (3) LINE : optimizes the objective function that preserves both the local and global network structures with an edge-sampling algorithm. (4) Driving State Vector (DSV) : the traditional transportation approach.
Evaluation 29 ¨ Evaluation Metrics v Square Error - Measure regression errors v Coefficient of Determination - measure the regression accuracy v Normalized Discounted Cumulative Gain(NDCG@N) - Evaluate the rankingperformance at TopN v Kendall’s Tau Coefficient(Tau) - Measure the overall ranking accuracy.
Overall performance 30 1.2 0.5 Auto − Encoder LINE Auto − Encoder LINE DeepWAalk DSV DeepWAalk DSV 1.0 CNN PTARL CNN PTARL Square Error 0.0 0.8 R 2 0.6 − 0.5 0.4 − 1.0 1.2 0.4 Auto − Encoder LINE Auto − Encoder LINE DeepWAalk DSV 1.0 DeepWAalk DSV 0.3 CNN PTARL CNN PTARL 0.8 0.2 NDCG@N 0.6 Tau 0.1 0.4 0.0 0.2 − 0.1 0.0 @5 @10 @15 @20
Robustness Check 31 Robustness check in the score-based group
Robustness Check 31 Robustness check in the driving-state-based group
Study of Peer and Temporal Dependencies 32 1.0 Auto − Encoder Auto − Encoder PTARL − temporal 0.5 PTARL − peer PTARL − peer PTARL PTARL − temporal 0.8 PTARL Square Error 0.0 R 2 0.6 − 0.5 0.4 0.2 − 1.0 0.4 1.2 Auto − Encoder Auto − Encoder PTARL − peer PTARL − peer 1.0 0.3 PTARL − temporal PTARL − temporal PTARL PTARL 0.8 0.2 NDCG@N Tau 0.6 0.1 0.4 0.0 0.2 − 0.1 0.0 @5 @10 @15 @20
Study of Performance in Different Views 33 0.7 Transition Probability View Transition Probability View Transition Duration View 0.5 Transition Duration View PTARL PTARL 0.6 Square Error 0.5 0.0 R 2 0.4 − 0.5 0.3 0.2 − 1.0 1.2 0.5 Transition Probability View Transition Probability View Transition Duration View Transition Duration View 1.0 0.4 PTARL PTARL 0.3 0.8 NDCG@N 0.2 0.6 Tau 0.1 0.4 0.0 0.2 − 0.1 0.0 @5 @10 @15 @20
Historical Assessment of Driving Scores 34 1.2 Riskier Driver Safer Driver 1.0 0.8 Score 0.6 0.4 0.2 0.0 5 10 15 20 Time
Risky Area Detection 35 t=2 t=1 t=3 t=4
Outline 35 ¨ Background and Motivation ¨ Definition and Problem Statement ¨ Methodology ¨ Application ¨ Evaluation ¨ Conclusion
Recommend
More recommend
