Overview / Usage

Driving involves different levels of scene understanding (perception) and decision making, lies in between detection and tracking of traffic participants, scene recognition, risk assessment, localization (where we are ) etc., to interact. To achieve an intelligent transportation system, we need a higher level understand.

First step is to detecting traffic participants and parsing scenes into the corresponding semantic categories.

To complete scene understanding, we need to understand the interactions between human driver behaviours and the corresponding traffic scene situations.

Towards intelligent transportation systems Learning is the first thing that how human drive and interact with traffic scenes. We need to start from a driver-centric view to describe driver behaviours.

Visual Scene recognition for autonomous driving required some datasets.

For example the dataset KITTI provides a suite for sensors and cameras also, Lidar and GPS/INS.

Ours dataset provides Driver behavior & causal reasoning for Camera, Lidar, GPS, IMU and CAN.

​

​

​

Methodology / Approach

We need to design Algorithm, architecture for our model. LSTM is the best choice for this task, these were shown to be successful in many temporal modelling tasks. LSTM process a video stream , our model has access to an auxiliary signal which provides complimentary info about the vehicle dynamics. This auxiliary signal includes CAN bus sensors: Car Speed, Accelerator and braking pedal positions, yaw rate, steering wheel angle, and the rotation speed of the steering wheel.

​

Technologies Used

Softwares / Architectures/ Tools

CNN (Conv 2d)

RNN (LSTM)

Tensorflow

Theano

Keras

Python

Pandas

Driving Datasets:

​

Other links

• Deep Residual Learning for Image Recognition
• Deep learning algorithms for scene understanding - MASK R-CNN
• Towards Real-Time Object Detection with Region Proposal Networks