Abstract
In this thesis, we study the multiple object detection algorithms and propose improved methods. The goal is to increase detection rate and reduce memory cost in multiple object detection methods. For the multiple object probability boosting tree scheme, we use tree node index to reduce memory cost and use boosted cascade to achieve high detection rate and low false positive rate. Furthermore, in multiple object capable cascade method, we propose the single stage cascade to replace the original parallel cascade structure to reduce the number of weak classifier and achieve high accuracy. For MIT CBCL car database, the multiple object capable cascades algorithm with single stage cascade structure reduces low memory cost by around 13.161x103 ~ 261.477x103 bits and achieves high detection rate of 95.54% and low false positive rate of 1.94% compared to above other algorithms. In the hardware implementation of multiple object boosted cascade scheme, we design an efficient the integral window to reduce the memory cost and calculate feature value. In cascade detection, we reduce the number of weak classifiers storage, and we exploit the parallel cascade detection architecture to reduce the detection time. The post-layout chip achieves operation frequency of 100MHz, processing images of 30 fps with size 160x120, and with core area of 1.21 mm2.