EEG based brain-computer interface (BCI) allows people to communicate and control external devices using brain signals. The application of BCI ranges from assisting in disabilities to interaction in a virtual reality environment by detecting user intent from EEG signals. The major problem lies in correctly classifying the EEG signals to issue a command with minimal requirement of pre-processing and resources. To overcome these problems, we have proposed, BCINet, a novel optimized convolution neural network model. We have evaluated the BCINet over two EEG based BCI datasets collected in mobile brain/body imaging (MoBI) settings. BCINet significantly outperforms the classification for two datasets with up to 20% increase in accuracy while fewer than 75% trainable parameters. Such a model with improved performance while less requirement of computation resources opens the possibilities for the development of several real-world BCI applications with high performance.