Privacy Preserving Multi-target Tracking Anton Milan Stefan Roth Konrad Schindler Mineichi Kudo
Visual People Tracking Applications and Benefits ✔ CCTV: Increased safety ✔ Automated video analysis ✔ Crowd motion estimation ✔ Robotic navigation A. Milan et al. | Privacy-Preserving Multi-Target Tracking 2
Visual People Tracking Applications and Benefits ✔ CCTV: Increased safety ✔ Automated video analysis ✔ Crowd motion estimation ✔ Robotic navigation Drawback: Heavy intrusion of privacy A. Milan et al. | Privacy-Preserving Multi-Target Tracking 3
Existing Alternatives [Schiff et al., 2009] [Wickramasuriya et al., 2005] [Spindler et al., 2006] A. Milan et al. | Privacy-Preserving Multi-Target Tracking 4
Existing Alternatives [Schiff et al., 2009] [Wickramasuriya et al., 2005] [Spindler et al., 2006] Such systems may fail (or be switched off) A. Milan et al. | Privacy-Preserving Multi-Target Tracking 5
Our Approach ● A different sensor modality ● Existing multi-target tracking techniques Pyroelectric infrared sensors * ... ...mounted on a ceiling * Also known as: Infrared motion sensors A. Milan et al. | Privacy-Preserving Multi-Target Tracking 6
The Setup 43 nodes, ca. 3m stride. Total cost: ≈ $100 USD. A. Milan et al. | Privacy-Preserving Multi-Target Tracking 7
Tracking with Infrared Sensors A mostly unexplored research area! [Luo et al., 2009] [Hosokawa et al., 2009] - Expensive sensor array - Limited state space - Ad hoc algorithm for data association with Fresnel lenses A. Milan et al. | Privacy-Preserving Multi-Target Tracking 8
Benefits ● Individal identification impossible – Respects privacy ● Insensitive to lighting conditions ● Low cost Limitations ● Indoor application only ● Less flexible A. Milan et al. | Privacy-Preserving Multi-Target Tracking 9
Main Challenges ● Extremely low resolution (43 sensors) ● A binary response at 2 Hz per sensor ● No visual (appearance) information ● Poor localization + sensor noise / delay ● Activation by several people ● Multiple measurements by one person A. Milan et al. | Privacy-Preserving Multi-Target Tracking 10
Main Challenges ● Extremely low resolution (43 sensors) ● A binary response at 2 Hz per sensor ● No visual (appearance) information ● Poor localization + sensor noise / delay ● Activation by several people ● Multiple measurements by one person A. Milan et al. | Privacy-Preserving Multi-Target Tracking 11
Main Challenges ● Extremely low resolution (43 sensors) ● A binary response at 2 Hz per sensor ● No visual (appearance) information ● Poor localization + sensor noise / delay ● Activation by multiple people ● Multiple measurements by one person A. Milan et al. | Privacy-Preserving Multi-Target Tracking 12
Main Challenges ● Extremely low resolution (43 sensors) ● A binary response at 2 Hz per sensor ● No visual (appearance) information ● Poor localization + sensor noise / delay ● Activation by several people ● Multiple measurements by one person A. Milan et al. | Privacy-Preserving Multi-Target Tracking 13
Continuous Energy Minimization E X = E obs E dyn E exc E per E reg X ∈ℝ d ,d ≈ 2000 State vector: X,Y -locations of all targets at all frames [Milan et al., PAMI 2014] A. Milan et al. | Privacy-Preserving Multi-Target Tracking 14
Why Continuous Energy? ● Continuous trajectories – low sensor resolution not an issue ● No implicit data association – multiple assignments possible ● Provides best results – Measured by standard tracking metrics A. Milan et al. | Privacy-Preserving Multi-Target Tracking 15
The Energy E = E obs + aE dyn + bE exc + cE per + dE reg physically-based priors data regularizer dynamics exclusion persistence parsimony − ∑ N + ∑ i 1 / length i ∑ ∑ − 1 − 2 ∑ i 1 exp 1 − b X i ∣ ∣ X i − D g ∣ ∣ − 2 t − v i t 1 ∣ 2 ∣ ∣ X i − X j ∣ ∣ ∣ ∣ v i ∣ g i ≠ j i A. Milan et al. | Privacy-Preserving Multi-Target Tracking 16
Data Term lobe size E A. Milan et al. | Privacy-Preserving Multi-Target Tracking 17
Optimization E( X ) Conjugate gradient descent Merge – Split Grow – Shrink Add – Remove Jump moves X ● conjugate gradient descent for local optimization ● discontinuous jumps to determine dimensionality (number of targets) A. Milan et al. | Privacy-Preserving Multi-Target Tracking 18
Experiments Synthetic Data ● Manual assignment of keyframes ● Interpolation and sensor simulation A. Milan et al. | Privacy-Preserving Multi-Target Tracking 19
Measurements A. Milan et al. | Privacy-Preserving Multi-Target Tracking 20
Measurements Time A. Milan et al. | Privacy-Preserving Multi-Target Tracking 21
Ground Truth Time A. Milan et al. | Privacy-Preserving Multi-Target Tracking 22
Experiments Synthetic Data ● Manual assignment of keyframes ● Interpolation and sensor simulation Result GT A. Milan et al. | Privacy-Preserving Multi-Target Tracking 23
Result Time A. Milan et al. | Privacy-Preserving Multi-Target Tracking 24
Result A. Milan et al. | Privacy-Preserving Multi-Target Tracking 25
Real Data Up to six people in ● a large lab Two cameras ● (2 Hz) Temporal ● alignment Annotation of key ● frames (very approximate) A. Milan et al. | Privacy-Preserving Multi-Target Tracking 26
Real Data Up to six people in ● a large lab Two cameras ● (2 Hz) Temporal ● alignment Annotation of key ● frames (very approximate) A. Milan et al. | Privacy-Preserving Multi-Target Tracking 27
Results (real) Result GT A. Milan et al. | Privacy-Preserving Multi-Target Tracking 28
Results (real) Result GT A. Milan et al. | Privacy-Preserving Multi-Target Tracking 29
Other Approaches [Tao et al., Sensors 2012] [Berclaz et al., PAMI 2011] A. Milan et al. | Privacy-Preserving Multi-Target Tracking 30
Quantitative Results MOTA = normalized error count Dataset Method MOTA [%] MOTP [%] ID sw #Targets (MAE) 13 Ours 76.0 73.6 0.54 synthetic MOTP = localization error (73% ≈ 35 cm) Ours 55.3 54.6 43 0.76 Real data A. Milan et al. | Privacy-Preserving Multi-Target Tracking 31
Quantitative Results Dataset Method MOTA [%] MOTP [%] ID sw #Targets (MAE) 13 Ours 76.0 73.6 0.54 synthetic Linear DA [1] 66.6 64.6 58 0.57 DP [2] 55.9 65.3 57 0.62 KSP [3] 75.5 67.5 6 1.52 Ours 55.3 54.6 43 0.76 Real data Linear DA [1] 9.3 50.1 252 1.00 DP [2] 9.6 47.3 128 1.25 KSP [3] 31.1 48.3 48 1.52 [1] Tao et al., Sensors 2012 [2] Pirsiavsah et al., CVPR 2011 [3] Berclaz et al., PAMI 2014 A. Milan et al. | Privacy-Preserving Multi-Target Tracking 32
Advertisement A. Milan et al. | Privacy-Preserving Multi-Target Tracking 33
Advertisement ● 22 Sequences (old + new) ● > 1300 Trajectories ● > 100,000 Bounding boxes ● Live online evaluation A. Milan et al. | Privacy-Preserving Multi-Target Tracking 34
Advertisement A. Milan et al. | Privacy-Preserving Multi-Target Tracking 35
Advertisement http://motchallenge.net A. Milan et al. | Privacy-Preserving Multi-Target Tracking 36
Summary ● A principled alternative to preserve privacy ● Continuous energy with soft assignments ● Still a very challenging problem ● Data + Code online http://research.milanton.net/irtracking/ A. Milan et al. | Privacy-Preserving Multi-Target Tracking 37
Recommend
More recommend
