ProjAEB_4thYear

ProjAEB_4thYear copied to clipboard

Project: Development of AEB System for Pedestrian Protection (4th year)

    funded by Ministry of Trade, Industry and Energy of Korea. (MOTIE)(No.10044775)

Goals

0. Research trend of sensor fusion for object detection
   • State-of-the-art
     • 3D Object Proposals for Accurate Object Class Detection [Chen, NIPS 15] [Project] [Code]
     • Multiview random forest of local experts combining rgb and lidar data for pedestrian detection [Gonzalez, IV '15]
     • Voting for Voting in Online Point Cloud Object Detection [Wang, RSS '15] [Project]
     • Pedestrian Detection Combining RGB and Dense LIDAR Data [Premebida, IROS '14] [Project] [Code]
     • Vehicle Detection from 3D Lidar Using Fully Convolutional Network [Li, RSS '16]
     • Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes [Yebes, Sensors '15]
   • Sensor configuration of autonomous vehicles
     • Google
     • Tesla
     • BMW
1. Build a testbed / Select datasets
   • Datasets: Traffic scenes
     • KITTI [Link]
       • Stereo, Lidar, GPS
       • Classes: Car, Pedestrian, Cyclist
       • GT: Bounding box
     • Cityscapes [Link]
       • Stereo, Timestamp
       • Groups: flat, human, vehicle, construction, object, nature, sky, void
       • GT: Dense pixel-level annotations
     • Virtual KITTI [Link]
       • Mono (forward / 15-deg-right, 15-deg-left)
       • Classes: Car, Pedestrian, Cyclist
       • GT: Bounding box, Instance-level pixel annotations, Optical-flow, Depth
       • Weather conditions: morning, sunset, overcast, fog, rain
     • Synthia [Link]
       • 8 RGB (form binocular 360 deg), 8 depth sensors
       • Classes: misc, sky, building, road, sidewalk, fence, vegetation, pole, car, sign, pedestrian, cyclist, lanemarking
       • GT: Instance-level pixel annotations
       • Seasons: winter, fall, spring, summer
       • Lightings: dynamic light, shadows, day-time, rain, night-time
2. Implement several algorithms
3. Evaluation / Comparison / Analysis

Related papers.

0. Unsupervised Depth Estimation. [Garg, ECCV '16]
1. LIDAR point upsampling. [Schneider, Arxiv '16]
2. Unified multi-scale CNN. (KITTI: 8th car, 1st ped) [Cai, ECCV '16] [Home] [Code] [Video]
3. Subcategory-aware CNN. (KITTI: 7th car, 3rd ped)) [Xiang, Arxiv '16] [Home]
4. Exploit all layers. (KITTI: 10th car, 5th ped) [Yang, CVPR '16] [Home]
5. 2D/3D Sensor Exploitation and Fusion for Enhanced Object Detection (Similar to ours) [Xu, CVPRW '14]

Why do I use github for this project?

See this. (Korean)

Easy & Pretty Documentation with Markdown

See this. (Korean)