Realtime feature extraction using MAX-like convolutional network for human posture recognition

This paper presents a realtime feature extraction processor based on MAX-like convolutional network. Due to the massive parallel MAX operations across multiple layers of feature maps, conventional implementation requires a vast amount of memory access as well as computation circuits. By exploring the overlapped data and reusing the intermediate computation results between consecutive "neurons", tremendous saving in both memory bandwidth and hardware resource has been achieved. Experimental results show that the number of logic gates drops from 402k to 170k, compared to conventional approach. The proposed feature extraction processor can be integrated with a custom-designed motion detection image sensor and a hardware-accelerated classifier to perform realtime human posture recognition.