Modified version of CNN-based object detection for omnidirectional vision sensors

点击翻译按钮获取中文摘要

Thales Company is currently investigating in developing low-cost optimal situational awareness systems for security. The current system is based on multiple vision sensors (visible and thermal wavebands) to detect threating objects under variant and difficult weather conditions. The sensors are manufactured with camera optics (fisheye lenses) that allow capturing a wide field-of-view extending 180-degrees to provide panoramic images and total coverage of the environment with lighter load. The main problem in these sensors is the optical distortion artifacts produced by fisheye lenses.The aim of this project is to explore a flexible and low-cost software solution independent of the optical model of the camera. The solution is not limited to adapting image processing algorithms to correct the distortion. The company is looking for an optimal solution that helps in reducing the inherent cost and power consumptions of the existing situational awareness system. An optimal software solution should be able to relax the distortion constraints in fisheye cameras and retain the effectiveness of the current CNN-based object detection algorithm compared to the standard cameras. A novel approach needs to be developed and implemented in this project with the objective of increasing the predictive capabilities of the CNN-based algorithm in less computation resources. The success of this project will allow Thales Canada to effectively develop and deploy low-cost security systems with minimal load and power across Canada and the world. Moreover, the success of this project will be of great impact from economic and environmental perspectives. Reducing the energy power consumption inherent in the current system implies less consumption of natural resources.