Convolutional Neural Networks (CNNs) are computational-intensive and resource-consuming. To build CNNs with low resource requirements of embedded computer vision applications such as object detection, we propose quad-approx networks. Although binarized networks are good for classification tasks they are not adequate for object detection. In quad-approx networks, we first quantize the convolutional layers. Features and weights for convolution are encoded into 3 bits. On top of that, an approximate multiplier for this special quantized network is proposed. Both approximations are back annotated to the training process leading to no loss in overall precision. The hardware simulation and experimental results are presented for quad-approx CNN based on Zynq UltraScale+ MPSoC ZCU102. 5.3x compression of network and 1.20x speedup for calculation are achieved.