AI-Video Pre-Processing Achieves Impressive Results for ADAS and AD Applications

Image

Introduction

Video processing has become a crucial functionality in modern cars. The accurate detection and interpretation of information from video streams is core to advanced driver’s assistance systems (ADAS) and autonomous driving (AD) applications.

By correctly perceiving and interpreting video streams from cameras, the vehicle’s software performing decisions - offered by computer vision - can increase the overall safety of the car. Therefore, understanding camera signals through intelligent software represents the cornerstone of ADAS (L2) and of fully autonomous vehicles (L4 and L5 level) and the advancement of autonomous mobility in general.

Modern cars can integrate up to 12 or more cameras (with different types of lenses, angles, and purposes) to gain a precise perception of the road and traffic participants interacting in real time. However, the amount of data generated in each of these cameras can amount to hundreds of megabytes, putting additional pressure on already constrained processors. Furthermore, most ADAS and AD use cases require a certain level of quality to ensure that the decisions executed by the software are reliable.

As cameras become the eyes of the future, it also carries challenges concerning hardware and software configurations, which could result in tricky pitfalls when interpreting video streams at the embedded level.

AI-Video processing achieves groundbreaking results

To overcome these issues, TERAKI utilizes artificial intelligence to understand the context of any given video, based on previously input data discerning regions of interest (ROI) and regions of no interest (RONI) in a camera frame. Following this approach, TERAKI has attained impressive results, achieving up to 76% in data savings using strong compression and 65% using lower compression compared to videos encoded with standard H.264 parameters.

The software developed by TERAKI consists of two steps: First, a detector which splits the camera frame into ROI and RONI (shown in red in the image below). Second, the encoder compresses the ROI and RONI differently with more aggressive compression applied to the RONI.

The red area shows the RONI areas detected by TERAKI’s software, which are heavily compressed.

The red area shows the RONI areas detected by TERAKI’s software, which are heavily compressed.

During the encoding process, the ROI areas are preserved in high quality, while the RONI areas can be compressed aggressively, therefore realizing added compression as compared to a standard H.264 encoder.

Employing this intelligent video compression software saves data transmission costs and preserves the quality of video streams. By reducing the size of the video files, the embedded architecture of the processors improves, releasing computational capacities to execute other functions, which is ideal for ADAS, AD and in-vehicle infotainment (IVI) systems.

TERAKI’s approach follows a pre-processing logic that ensures the highest video quality for less data usage.

TERAKI’s approach follows a pre-processing logic that ensures the highest video quality for less data usage.

The results have been proven on an NVIDIA Jetson Nano, a SoC with comparable computational power as used in automotive applications.

Makes for better detection and classification

Applying a ROI/RONI approach to video applications not only improves the resulting quality for IVI systems, but also leads to enhanced AI performances for other L2+ applications. For the automotive industry, passenger security and road safety are top priorities bound to the performance and accuracy of machine learning models. In this sense, ADAS and AD features require model training and testing, whose precision leans on the data quality fed into subsequent machine learning models.

TERAKI’s pre-processing addresses these data demands by both reducing data (hence, improving latencies) while preserving image quality – which ensures the correct identification of traffic participants. The ROI/RONI models improve the reliability and accuracy of the OEMs L2+ models for use cases such as lane departure, pedestrian detection/avoidance, traffic sign recognition, and more.

On the left, compression results from regular video encoders result in poor video quality. On the right, TERAKI encoders achieves the same data compression with remarkable video quality.

On the left, compression results from regular video encoders result in poor video quality. On the right, TERAKI encoders achieves the same data compression with remarkable video quality.

Benchmarking with real data recurrently confirmed 20%-30% better F1 scores (measurement for AI-accuracy). This means that the car’s L2, L3 and L4 software becomes more precise and safer. For OEMs, the results translate into safer – hence better – cars, as it could potentially reduce the number of accidents

Advancing automotive technology through intelligent AI software

TERAKI’s mission is to enable safe, cost-effective autonomous driving applications through cutting-edge software. By efficiently processing video signals, TERAKI’s video preprocessing improves and advances ADAS and AD applications in the automotive industry.

Contact us at info@teraki.com to learn more about our automotive solutions and sign up for our newsletter to stay informed of our latest developments.

Share this Post: