A cognitive framework for object recognition with application to autonomous vehicles

Abstract

Autonomous vehicles or self-driving cars are capable of sensing the surrounding environment so they can navigate roads without human input. Decisions are constantly made on sensing, mapping and driving policy using machine learning techniques. Deep Learning – massive neural networks that utilize the power of parallel processing – has become a popular choice for addressing the complexities of real time decision making. This method of machine learning has been shown to outperform alternative solutions in multiple domains, and has an architecture that can be adapted to new problems with relative ease. To harness the power of Deep Learning, it is necessary to have large amounts of training data that are representative of all possible situations the system will face. To successfully implement situational awareness in driverless vehicles, it is not possible to exhaust all possible training examples. An alternative method is to apply cognitive approaches to perception, for situations the autonomous vehicles will face. Cognitive approaches to perception work by mimicking the process of human intelligence – thereby permitting a machine to react to situations it has not previously experienced. This paper proposes a novel cognitive approach for object recognition. The proposed cognitive object recognition algorithm, referred to as Recognition by Components, is inspired by the psychological studies pertaining to early childhood development. The algorithm works by breaking down images into a series of primitive forms such as square, triangle, circle or rectangle and memory based aggregation to identify objects. Experimental results suggest that Recognition by Component algorithm performs significantly better than algorithms that require large amounts of training data.