1 of 11

Fast Obstacle Detection Using StixelNet

Shahar Zuler & Noa Raindel

03.06.2018

2 of 11

Stixelnet - Fast and Efficient Obstacle Detection

What is the minimum required for obstacle detection?

    • Recognition?
    • Detection of all objects in the frame?
    • Where is the closest obstacle. No pixel or patch level.

​

Levi et al., 2015

3 of 11

KITTI Dataset Overview

  • Images and depth information captured from a driving car.
  • Ground truth for obstacle location was provided by Dan Levi.
  • After pre-processing:
    • ~1M stixels
    • ​1⁄3 have ground truth data

​

​

http://www.cvlibs.net/datasets/kitti/index.php

4 of 11

The Main Goal is to Implement a Faster Version of StixelNet

LeNet

(StixelNet)

MobileNet V2 Basic unit

MobileNet V1 Basic unit

Levi et al., 2015

Sandler et al., 2018

5 of 11

Performance Metrics Include the Accuracy (Top 1) and Area Under Curve (AUC)

6 of 11

MobileNets Show Similar Performance While Being Computationally Lighter Then StixelNet

  • The paper used LeNet architecture
    • Slow, 3 fully-connected layers, computationally expensive
  • MobileNet is a fully-convolutional network
    • Fast, fewer parameters
  • MobileNet_V2 has residuals thus achieves better results
  • Overfitting small dataset:
    • Very fast on MobileNet_V1 and MobileNet_V2

​

​

Architecture

# parameters

Top-1 accuracy

AUC

LeNet (StixelNet)

20.7 M

49.7% *

0.86

MobileNet V1

4.9 M

37.6%

0.844

MobileNet V2

1.7 M

40.7%

0.851

MobileNet V2 asymmetric kernels

1.8 M

39.5%

​

0.849

* from Levi et al., 2015

7 of 11

MobileNet performances are similar to StixelNet

  • Changing kernels size (H>W) improved results
  • Using MobileNet_V2 (has residuals)

​

​

​

​

Levi et al., 2015

LeNet (StixelNet)

MobileNet

8 of 11

Output examples from MobileNet show satisfying results

Our prediction

Ground truth

9 of 11

Things We Learned

  • Real life data is “Dirty”.
  • Built pipeline with Dataset and Estimator APIs.
  • All MobileNet architectures showed similar results to StixelNet.
  • Residual blocks architecture (MobileNet V2) is more efficient.
  • Rectangle kernels did not improve accuracy.
  • Higher number of steps might lead to�overfitting on train data.

​

10 of 11

Future work suggestions

  • Better generalization:
    • Add dropout
  • Data augmentation

​

  • Solve as regression problem

​

  • Fine-tune hyperparameters

​

​

11 of 11

Thank you!

Nexar:

Shmuel Rippa

Elad Levi

Ilan Kadar

Course Staff:

Uri Eliabayev

Eran Paz

Gil Levi

Nir Ben-Zvi