Multi camera tracking. Jan Baan. Content. VBM Multicamera VBM A270 test site Helmond

Size: px

Start display at page:

Download "Multi camera tracking. Jan Baan. Content. VBM Multicamera VBM A270 test site Helmond"

Bridget Skinner
5 years ago
Views:

1 Jan Baan Content VBM Multicamera VBM A270 test site Helmond 2 1

2 3 Introductie/Doel video vehicle detection image to world tracking video track filtering VBM Database Vehicle data trajectory (x,y,t) dimensions type datebase 4 2

3 Vehicle detection, training examples no vehicles (negatives) train video > 1000 examples (positives) DETECTOR 5 Vehicle detection examples no vehicles (negatives) train video > 1000 examples (positives) DETECTOR 6 3

4 Vehicle detection DETECTOR video detections 7 Image coordinates to world coordinates Camera calibration World, Detection in meters Image, detections in pixels 8 4

5 9 Tracking time single detections 10 5

6 Tracking time single detections tracks 11 VBM characteristics meter (with no occlusion) 97%-99% of vehicles detected Processing time off-line High Performance Cluster of 200 nodes (~50 quad core PCs) > 20x realtime Processing realtime on a PC Database of 300 days Sun, Rain, Night, Snow No ground truth 12 6

7 VBM Multicamera Tracking 700 m 600 m Cam m 4 Cam m Cam m m Cam m m 1 Method multicamera algorithm Peel method used 1. Overlap => Tracks removed from set 2. Visual features (SIFT) => Tracks from set 3. Extrapolation 14 7

8 Last step extrapolation: global optimization Cluster tracks, more possibilities for linking Two methods 1. Connect first best fitting tracks 2. Global optimization (better) Cluster tracks First connect best fitting tracks Global optimization 15 Multicamera Tracking 700 m 600 m Cam m Cam m Cam m m Cam m 2 16 Video Multi camera Based traffic tracking Monitoring (VBM) Woensdag dinsdag november m 1 8

9 VBM Multicamera results Multicamera VBM Distance between cameras 100 meter: > 99% vehicles correct connected > 99% vehicles connected Distance between cameras 200 meter: 98% vehicles correct connected 96% vehicles connected 17 DEMO 18 9

afterwards Application: Shockwave elimination/reduction Red/left

10 A270 Helmond-Eindhoven, 20 camera s 19 A270 test site First experiments in February camera s VBM used for evaluation afterwards Application: Shockwave elimination/reduction Red/left vehicles: not instrumented Grey/right vehicles: instrumented with speed advise device 20 10

11 Shockwave elimination/reduction 21 A270 Test site Helmond March camera s, 4.8 km, each 100 meter a camera Server room with 20 high end PCs Live tracking Applications Shockwave reduction/elimination Merging assistant Cooperative driving VBM is used as sensor Integration other sensors Radar Lidar FCD/OBU VBM detector server VBM detector server VBM detector server VBM detector server Tracking server OBU OBU Service Platform FCD/OBU data OBU 22 11

12 Datafusion/tracker All detections in same coordinate system Trackertoolbox Track prediction Association of detections to tracks Detection position uncertainty Update track, kalman filtering Xnew = α Xdet + (1-α) Xpred X is a state: position vector and speed vector α dependent on sensor type data type (position or speed) position camera s 16 VBM detectors PCs 1 tracker PC Integrate Floating Car GPS Data Tracks are used for shockwave reduction algorithm First experiments March

13 A270 testsite 4.8 km, 48 camera s Helmond Eindhoven 25 Conclusions VBM meter (with no occlusion) 97-99% vehicles detected and tracked Multicamera VBM Distance between cameras 100 meter: > 99% vehicles correct connected > 99% vehicles connected Distance between cameras 200 meter: 98% vehicles correct connected 96% vehicles connected A270 test site Helmond (March 2011) 5 km, 48 camera s, each 100 meter a camera Live processing Integration other sensors 26 13

Realtime Object Detection and Segmentation for HD Mapping

Realtime Object Detection and Segmentation for HD Mapping William Raveane Lead AI Engineer Bahram Yoosefizonooz Technical Director NavInfo Europe Advanced Research Lab Presented at GTC Europe 2018 AI in