In this study, we develop a multitasking CNN (Convolutional Neural Network) for advanced driving assistance. The network simultaneously performs three tasks: object detection, semantic segmentation, and disparity estimation. Edge computing requires low computation and low storage capacity, so the three tasks share not only one encoder, but also one decoder that employs a combination of depth-wise point-wise convolution and bilinear interpolation instead of the usual transpose convolution. This reduces the number of multiply-accumulate operations to 44.0% and the number of convolution weight parameters to 38.2%. In multitasking CNN training, the loss weights for each task were automatically adjusted by backpropagation, and the three tasks were learned in a balanced manner. Reducing the complexity of the decoder did not degrade the recognition accuracy, but rather improved it. Moreover, we found that entering pixel coordinates in this CNN significantly reduced misestimations for images that differed significantly from those during training.