ABOUT NORCOM
Deep learning for autonomous driving
The task
NorCom developed an advanced analytics & machine learning workflow for an automobile manufacturer. The aim was to investigate the dependence of driver behavior on the driving situation.
The challenge
The sheer amount of data, whereby time series with image data in the double-digit terabyte range have to be merged for evaluation on GPUs.
our solution
Relevant events (eg "ADC will be switched off") were filtered from the sensor data using a fast, signal-based search and synchronized with the images simultaneously recorded by the on-board camera. A deep neural network classified the situation recognizable in the images (eg "truck ahead"). Then statistics were created about the events (eg “distance to the truck”) and made comparable about situations.
The customer benefit
The consistent use of big data to avoid data movements together with the modular structure ensures the performance of the workflow even on mass data. Both the parameters and the algorithm of the workflow can be easily adapted for the customer so that the results can be interpreted iteratively with further data analyzes. Instead of a single finished analysis, the customer receives a template that he can apply to many other use cases.
Project-
Characteristics
Our role
Customer support from data scientists, data engineers, architects and software developers
Our activities
Setting up yarn queues for the allocation of GPUs
Preprocessing of measurement data
Creation of big data capable deep learning workflows
Training in the use of Tensorflow
Technologies & methods
Applications: DaSense
Databases: Hbase, Hive, HDFS, ADTF, Parquet, ROS
Languages / Frameworks: Python (Anaconda Stack), Java, Java-script, TensorflowOnSpark, Hadoop / MapR, Spark, Yarn, Oozie
Methods: time series analysis, image analysis, sensor fusion / synchronization, deep learning