A reduced-complexity, reduced-power camera system for intrusion classification in an outdoor setting

An optical camera was recently employed by a subset of the authors as a sensing modality complementary to that of a Pyroelectric InfraRed (PIR) sensor, for carrying out intrusion detection and classification in an outdoor environment. The aim there was to develop a classification algorithm that mimicked the performance of the PIR sensor and which was complementary to the PIR in the sense that it could carry out detection in environments where the PIR was unable to function effectively. An example of such a setting is when the ambient temperature is close to that of the human body. The algorithm for intrusion detection using the optical camera presented there, achieved an accuracy in excess of 95% and was implemented on an Odroid C2 (OC2) single-board computer (SBC). However, at 3.6W, the OC2 is relatively power hungry and this prompted the current effort to develop a version of the optical camera system that could be operated at a reduced power level. The power-reduction techniques considered here include examining alternative SBCs such as the Orange Pi Zero (OPiZ), developing reduced-complexity intrusion-detection algorithms that permit a reduction in power consumed through a commensurate reduction in CPU clock speed, operating with a single CPU core and reducing power consumption by peripherals. Reduction in computational complexity was achieved by working with a reduced image resolution, a smaller Lucas-Kanade window, identifying and deleting power-consuming steps that did not contribute appreciably to classifier performance as well as removing portions of the code used for understanding and debugging. The resultant reduced-complexity algorithm, implemented on an OPiZ, turned in performance comparable to that of the high-complexity algorithm at a reduced power consumption of 1.3W. In most instances, classification accuracies obtained on recorded data were in excess of 95%. An outdoor deployment was also carried out in which 1.23% missed intrusion detections, 2.22% intrusion misclassifications, and 1.72% false alarms were observed.