Reliability evaluation and enhancement of stereo vision accelerators
Parole chiave FAULT DETECTION AND IDENTIFICATION, FAULT INJECTION, FAULT TOLERANCE
Riferimenti MATTEO SONZA REORDA
Riferimenti esterni Juan David Guerrero
Gruppi di ricerca DAUIN - GR-05 - ELECTRONIC CAD & RELIABILITY GROUP - CAD
Tipo tesi SPERIMENTALE
Descrizione Nowadays, vision-based systems are crucial in many cutting-edge applications, especially in the field of autonomous systems such for example robotics, automotive, and health care. These vision systems offer any artificial platform the ability to "see" the surrounding environment and make decisions to navigate in unknown environments performing specific tasks autonomously, all of this only relying on a couple of cameras. Autonomous driving systems are a clear example of the success of the vision technology since it gives automobiles the capabilities to navigate through complex routes reducing the stress of the driver as well as helping to improve the traffic conditions.
Stereo vision is one of the most important algorithms for autonomous systems since it calculates the depth of each object present in the visual range of the cameras. This information is crucial to detect obstacles helping the systems (e.g., automobile, robot, or drone) to create alternative routes in order to accomplish their assigned task successfully. However, the stereo vision resorts to complex algorithms which demand high computational power to acquire and process the images in real-time scenarios. Therefore, the incorporation of specialized hardware accelerators for stereo matching is essential to fulfilling the operational requirements of the application and maintaining the accuracy of the depth calculations.
There are many stereo-matching algorithms and hardware accelerator architectures proposed in the literature, such as Sum of Absolute Distances (SAD), Sum of Squared Differences (SSD), Normalized Cross Correlation (NCC), rank, and Census, among others. Each correlation method has advantages and disadvantages, but the census method provides the highest correlation accuracy and is much more robust than other local stereo methods w.r.t common image defects. On the other hand, the census approach is also more expensive to implement physically than the SAD or rank approaches due to the large number of bits required by the census window (e.g., 48 bits for a 7X7 census) requiring extra hardware for the calculation of the Sum of Hamming Distances (SHD). Nevertheless, the census transform plus the SHD has some redundancy properties, which favor the use of sparsity techniques reducing the number of pixels to be compared and significantly reducing the hardware resources without impacting the stereo matching accuracy.
Although the design efficiency of stereo matching accelerators is essential in terms of power consumption and performance, reliability is also a crucial aspect to be considered, especially in safety-critical applications like autonomous driving systems. Since any hardware failure in the stereo matching accelerator may induce noisy depth map images or, in worse cases, produce wrong object distance calculations giving incorrect information to the system and leading to catastrophic results. Therefore, it is necessary to devise strategies to enhance the reliability of these important hardware accelerators guaranteeing the requirements of safety standards.
This proposal aims for three main goals: i) devise methodologies to evaluate the criticality of the failures in the device, ii) propose test techniques for fault detection during in-field operation, and iii) develop hardening strategies to counteract any possible failure that endangers the system's correct operation.
Since this is a research-oriented thesis (with significant practical impact), the real activities will depend on the evolution of the project. Since a thesis corresponds to 30 credits, the student is expected to work on it for about 6 months. The evaluation of the performed work will be based on his/her commitment and skills, rather than the (difficult to forecast) achieved results.
Conoscenze richieste Digital Design
Scadenza validita proposta 21/09/2023 PROPONI LA TUA CANDIDATURA