The paper describes a new object detection algorithm which uses a convolutional network with a convolution kernel of the Network in Network (NiN) type. Detection refers to the simultaneous localization of objects on an image and their recognition. Detector can operate with images of arbitrary sizes. The algorithm proposed has a high computational efficiency, so when processing HD frame on a single CPU core, the operating time is about 300 ms. High degree of uniformity of network operations allows to massively parallel data processing on GPU, which is likely to reduce the operating time to less than 10 ms. Our algorithm is robust to small overlaps and the average quality of images of detected objects. It represents an end-to-end learner model, which output is delimited by boundaries and classes of objects throughout an image. An open access image database from car recorders is used to evaluate the algorithm for object detection. This algorithm is not limited to the use of one type of objects, it is able to simultaneously detect a mixture of objects.