The Use of Neural Network Techniques to Identify Small Objects for Self-driving Cars
Transportation has accelerated the process of human transformation from hunters and gatherers to a powerful society that is on the threshold of AI revolution. The advent of big data capabilities and exponential increase in computational power allowed for the paradigm shift in transportation. Humanity is about to enter the world of autonomous vehicles (AV) that are going to change completely the way people have moved around. With those changes come the new challenges. One potential problem that will arise is the inability of cars to detect small objects properly. Current sensors used in AVs are accurate when it comes to detecting big objects like cars, pedestrians or deer, however they fall short when it comes to objects like squirrels, trash and debris. These small objects can be misidentified by the AVs and trigger the system to take actions irrelevant to the situation. Those errors jeopardize the safety of AV riders.
This thesis proposes an object detection methodology which addresses the issue by training YOLO (You Only Look Once) real time detection architecture based on convolutional neural networks to identify custom classes like squirrels. This approach is transferable and can be used to identify any other desired class of objects. The data from object detection can be fused with the data from other AV’s sensors to ensure safe self-driving experience.