Object Detection and Identification for Car Reverse Based on Convolution Network Computation

This paper presented a computer vision method for obstacle detection combined with various algorithms. Deep learning methods are further used to recognize the obstacle. The system integrates with the framed obstacle detection and AI recognition system for vehicle systems. The advantage is that any obstacle can be detectable even if they are not trained. The obstacle is detected and put the target object into the waiting list linking to AI kernel. All images of the detected obstacle are passed to the AI recognition system at once. Each obstacle can be accuracy classification according to the training labels from the AI recognition. The results are shown on the display screen with the obstacle name and the distance. When the distance that is estimated with bird view method is less than one threshold, the system displays the obstacle warning sign. We collect the dataset with real photos for training database, which includes various public datasets and our lab self-build. The classification includes bicycle, motorcycle, dog, car, and pedestrian, where 1000 photos of each category. We can classify the common road categories with near 100% accuracy.