Abstract: With the continuous increase in high-rise buildings, incidents of objects being thrown from heights have also increased, posing challenges to personal and public safety. The detection of objects thrown from heights faces several issues, such as complex backgrounds, small targets, indistinct appearance features of the thrown objects, and difficulty in tracking these objects. To address these issues, the Mixture of Gaussians Background Modeling(MOG) algorithm has been extended using neural networks, and the Simple Online and Realtime Tracking(SORT) algorithm has been improved based on parabolic characteristics. First, regional conditional filtering is introduced to reduce non-target foreground in foreground detection, addressing the problems of small target objects and complex backgrounds. Second, multi-frame fusion technology is used to enhance motion features, and a lightweight classification network is designed to distinguish thrown objects, solving the problem of indistinct appearance features. Finally, the state space and matching metric of SORT are improved based on parabolic characteristics to address the difficulty in tracking thrown objects. Experimental results show that the improved MOG algorithm reduces detection numbers by 97.14% with a 6.50% decrease in recall rate; the improved SORT algorithm reduces ID switch numbers by 51.61%, improves the MOTA metric by 8.74%, and improves the TIOU metric by 8.02%.